This naming scheme is awful. As if the names weren't already confusing with Pro and Max, now we get Ultra, which is double Max. So okay, now your Maximum wasn't actually maximum, but half a maximum.
Yeah, how are we supposed to keep track of M1, M1 Pro, M1 Max and M1 Ultra. So confusing.
No no, give me Core i5 124800T, 12400F, 12400, 12490F, 12500T, 12500, 12600T, 12600, 12600KF, 12600K, Core i7 12700T, 12700F, 12700, 12700KF, 12700K and Core i9 12900T, 12900F, 12900, 12900KF and who can forget 12900K? And don't forget the mobile variants Core i5 12450H, 12500H, 12600H, Core i7 12650H, 12700H, 12800H, Core i9 12900H and 12900HK.
That's something we, the consumers, can get behind.
It's not foolishness, people (still) hold apple to higher standards, because they themselves have set them selves up like that. I mean it does not make too much sense though Like, how ironic is their "think different" campaign now.
One difficulty is that they hide different SKUs under the same name. At least with all those ridiculous variations, there *ARE* differences, however minor, between them all.
Apple has two M1 chips (7-core GPU and 8-core GPU,) 3 M1 Pro chips (8 or 10-core CPU, 14 or 16-core GPU, with 8+16 being the only disallowed,) 2 M1 Max chips (10-core CPU, 24 or 32-core GPU,) and now two M1 Ultra chips (20-core CPU, 48 or 64-core GPU.)
So nine different CPUs hiding behind only 4 names, with no way to tell the difference other than the longer-than-a-product-number actual description of CPU and GPU cores.
And we don't know about thermal envelope/speed differences between product lines, either. Does the M1 Max in the Mac Studio run faster than an M1 Max in a MacBook Pro? With Intel's product numbers, while they may be obtuse on their face, each one defines a specific spec of cores, speed, extra features, and power envelope.
I can assure you that Apple uses SKUs for all variations of their chips to keep track of their stock. Since they don't sell the chips to others they don't need to offer up confusing SKUs to customers of their machines.
Having a designator based on an architecture, scale, with an in-line numerical references to cores, RAM, and storage to be much easier to figure out without having to look up what a long coded SKU means.
As commented elsewhere, "Ultra" is Latin for "on the far side of, beyond", so it's not really incorrect. But I still agree these naming schemes are silly. Is the Mac Pro going to be the M1 Turbo^2?
But the naming is correct. Max means maximum which is the maximum die size. And Ultra is Latin for "beyond" which means they go beyond maximum die size by having two dies. Linguistically and logically the naming is rock solid.
"This means twice as many CPU cores, twice as many GPU cores, twice as many neural engine cores, twice as many LPDDR55 memory channels, and twice as much I/O for peripherals."
Really excited to hear about this LPDDR55! Skipping 50 generations is pretty impressive.
I think the assumption is that the 2.5TB/s Ultrafusion interconnect is higher bandwidth than the UCIe/PCIe5 bandwidth not that I can find information as the chiplet to chiplet bandwidth of UCI; is it PCIe5/CXL that determines the chiplet to chiplet bandwith?
They mention 1.3TB/mm, but not how many lanes (it says spare lanes in advanced), but I assume at least 2x, so 16 lanes?
A search reports the M1 Max getting 10MH/s. It's pretty power efficient using 19 extra watts, but good luck with reaching ROI... ever. Double that you get 20MH/s, which is what you'd get with 6 year old RX 470 at stock settings.
oh ? not going by the stock levels of a couple of local stores here, most cards are still out of stock. have havent changed for quite a while now. been keeping an eye on it as i would like to upgrade my GTX 1060 :-)
Mac Studios are way too expensive for non-creatives to try and use for use cases like mining at scale. It's far more sensible to get a 3090/6900XT until point-of-stake crypto-mining eventually
The performance claims of exceeding RTX 3090 is total nonsense. Maybe in some mobile GFXBench where it's optimized for mobile architectures. Maybe doubling M1 Max can get RTX 3080 Mobile performance.
Thus far most of Apple's performance claims about its silicon have borne out. Obviously independent testing is important, but IMO they've earned the benefit of the doubt.
That is because Apple has either: A. made general claims without referencing a specific device B. made specific claims concerning Apple devices running 8th, 9th and 10th gen Intel Core i3, i5, i7 and i9 CPUs. C. after it was demonstrated that Alder Lake - and some Tiger Lake and AMD - CPUs outperformed the M1, M1 Pro and M1 Max in configurations that cost hundreds less Apple subtly moved the goalposts to "industry leading power per watt" (in large part due to using TSMC's industry leading process that AMD was as much as 2 generations behind and Intel up to 4).
Apple has proven that they can make comparable CPUs to Intel and AMD but not better ones. They have a massive lead with integrated GPUs that Intel is actually starting an entire new research division to close the gap but that is the only area where they have a clear advantage. All the others are areas and metrics chosen by Apple for the purposes of giving them an advantage.
Finally, in addition to being on the latest TSMC process, a huge chunk of the M1's performance is due to replacing RAM with unified memory. Intel and AMD have no interest in copying this directly, but the chiplet idea that Intel has proposed and the rest of the industry (except Apple and Nvidia) has agreed to adopt has the potential to approximate it. It won't integrate the RAM on the SOC proper, but it will connect the RAM to the CPU on a connection that provides performance similar to what Apple is using to connect two M1 chips to make the M1 Ultra.
Apple has made performance per watt their primary goalposts since the announcement of M1, that was the first and most prominent slide. And I haven't seen Alder Lake outperforming the M1 Max in anything with a battery life over 4 hours.
We expect Alder Lake laptops to be cheaper because they'll be made in cheaper build quality laptops without premium design features from vendors stuck with WinPC 2% margins instead of Apples 20% margins. But I haven 't seen any yet.
I haven't checked in a while... however, if you wanted both the performance and the build quality (display quality, metal cases, ...) of the Apple, you usually paid the Apple price (but buying Dells, ...).
With regards to your comments: A. Actually, if you look at the footnotes of https://www.apple.com/mac-studio/">Apple's Mac Studio page they give the specs of the systems they benchmarked the M1 Max and M1 Ultra against. Ditto for the slides they used in the keynotes, the system specs are given on the bottom right. One of the systems they benchmark it against runs a Core i9-12900K, i. e. the fastest consumer-level Alderlake processor. B. And they do so for Alderlake, i. e. 12th gen Core processors, too. C. Performance-per-watt and its effect, battery life, were selling points from the very beginning. When the M1 Max was benchmarked, AT remarked that only a high-end desktop part with a much, *much* higher power consumption could beat it. Of course, time moves on and processors with much higher TDP have beaten M1-class processors in various benchmarks.
To claim Apple is “conveniently” switching to performance-per-watt as their yard stick is inaccurate. In fact, it was the reason they switched to Intel over 15 years ago. Intel knows it is seriously behind: in a leaked company slide deck, Intel employees were claiming that they intend to beat the M1's efficiency in 2024.
Lastly, you write that Apple's advantage is due to replacing RAM with unified memory. That's where some of its advantage comes from, yes. So what? This seems like a very good idea to me. Previously, the big downside of an integrated GPU was that memory bandwidth was scarce: CPUs used to have a much lower memory bandwidth to begin width, and the GPU gobbins had to share the meager bandwidth with the CPU. Hence, the association of integrated graphics = slow graphics. Once you lift the bandwidth bottleneck and allow for high-enough clockspeeds, you see that there is no performance compromise, but the opposite rather.
Yes and no. Apple has a history of relatively accurate performance claims. They are typically in line with a broader mix of benchmarks and have been verified here on Anandtech. In recent years that hasn’t always been the case with Intel. For this reason, I’d take that as a ballpark figure and am looking forward to the in-depth review here. :-)
And by ‘Day1’ you could easily mean day 1 of the transition from PPC to intel. It was IBM’s inability to fab a G5 that could go in a PowerBook that pushed apple to intel in the first place.
The M (and A) chips reflect apple priorities going back 20 years. In addition to performance/watt being a long-standing priority, specialized SIMD units (remember altivec?) and using the GPU performance for non-gaming purposes has also been a long time priority.
Basically apple had made the chip they always wanted and that intel could not make for them.
Yeah I think this is absolutely the take. Apple has made a huge part of its laptop segment since moving to intel about superior battery life. Only tangentially related, but I feel like they haven't pursued that on iPhone to nearly the same extent.
Yes and no on the iPhone. They have pursued power efficiency on their mobile devices, but more as a tradeoff for making those devices thinner. Generally, the battery capacity on iPhones is smaller than equivalent Android models, so the efficiencies with the SoC and OS optimizations have been deployed more in support of that tradeoff.
what are you talking about? they claim nearly 2x the performance of competitors even with competitors pumping up power draw, and better GPU performance than the 3080 laptop edition, yet they've never hit those numbers.
This Youtube video didn’t even run a proper benchmark. If you are going to make bold statements like it’s faster than the 3080 mobile in Tomb Raider, please actually back it up with a proper source where some semblance of a benchmark is conducted.
In the mean time, I would much rather rely on Anandtech’s results and Hardware Unboxed.
The Anandtech article you quote concludes that the M1 "chips here aren’t only able to outclass any competitor laptop design, but also competes against the best desktop systems out there"
no, i just watched a video of a 3080 mobile laptop get beaten by the M1 Max, it depends on the TDP of the GPU in that laptop of course, check out Tally Ho, the video link is above
Shadow of the Tomb Raider makes a poor comparison point, since it's running x86 code through Rosetta and using OpenGL instead of Metal.
World of Warcraft would be an example of a game that is both ARM native and targeted to Metal so at least you are seeing the full performance of the hardware.
Sort of like an how an x86 based XBox could run some of the older games written for PowerPC based XBox consoles, but emulation always slows things down.
I mean, most independent testers were unable to really confirm much of the claims of the M1 Pro vs M1 Max, specially because there is really not that many applications that make use of things like Metal 2 to really compare those claims.
The performance of the M1 Pro and M1 Max, is impressive in many factors, specially for battery life, but outside of that, they are still waaaaaaay too far away from high end desktop grade performance, not to mention that they are not even close to workstation/server grade CPUs. Maybe the Ultra will be able to finally close that gap with desktop grade performance, but, as nvidia well knows at least for GPUs, adding double the GPUs dies does increase performance by double, but also introduces double the problems to the point that Nvidia has mostly given up on that concept for a while now.
I think the fact that they have to pull out this trick of making 2 chips in order to finally do a desktop shows one of the main drawbacks of using ARM64 as an x86-64 replacement in this space. Lets just see how it goes, but I'm very skeptical of their wild claims.
Exceeding RTX 3090? Sure, total nonsense. RIVALING RTX 3090 ON AN INTEGRATED GPU? The opposite of total nonsense considering that Intel's competitor - the Iris Xe - doesn't even surpass an MX350, let alone an RTX 3090.
While I think RTX 3090 is the faster silicon, I wouldn't knock integrated GPU performance.
A properly designed integrated GPU would feature unified memory, which allows for zero-copy, and removes a lot of bandwidth bottlenecks inherent with Ram stick based CPUs joined with PCIe based GPUs.
PS5 is a good example of fast integrated GPU with unified memory. It's just that PS5's GPU is bound by cost constraints.
It doesn't matter if it's an Integrated GPU or not.
The amount of transistors they are dedicating to the GPU is more than what the 3090 uses so it makes sense.
The other difference, as Ryan states, is that since Apple goes extremely wide, they can jam all these transistors together and run them at low speeds so they use much less power! One of the big reasons that AMD can't put a 6900XT on a Ryzen 9 die is because they run high clock speeds on both and the heat would be unmanageable.
It's one hundred percent accurate, why can't you believe it. The computer costs $5000 for god's sake. The RTX 3090 is beaten by the $1000 6900 XT in comparison.
and nobody should use ray tracing, ray tracing is total crap, I play at 1440p and 120fps, and you can't do that with any modern game with ray tracing on
120 FPS is irrelevant for some genres of games that benefit from rich graphics, such as turn-based RPG. Those require few frames but benefit from graphical richness. Various simulation games do not require a high frame rate.
I find it humorous and unfortunate that the mentality of PC gamers continues to be locked into FPS-style games. It’s an entire genre of art; there is a lot more possible.
While I agree that so far the overall RT performance has not been up to snuff even with this generation of cards (particularly the 6900XT), there are always a portion of gamers who prefer eye candy to throwing 200fps at the screen. If there weren't, Nvidia and AMD wouldn't have bothered supporting that feature.
You seem stuck on the association integrated graphics = slow graphics. Current-gen consoles already proved this wrong, and finally with the M1 Pro and M1 Max we know that this also works for other platforms. Since Anandtech has already benchmarked the M1 Max, we know what to expect for the Ultra. In gaming benchmarks Apple's chips aren't doing so great, but that is down to drivers. In productivity applications the M1 Max (not the Ultra) already got into the same spheres as fast discrete desktop graphics. Plus, the RTX3090 is getting old and is built on an older process, so Apple's claims aren't surprising.
Being "integrated" is now an advantage. I fail to see how GPUs separated from the CPUs and memory are somehow going to be faster than when they are all bundled together using a common memory space.
Spoken like a true gamer. You do realize that the world is larger than games, right? A HUGE aspect of these chips is that they provide a much larger memory capacity for the GPU than either the 3090 (24G) or even a maxed out 48GB Quadro. Which is of immense importance to, for example, people playing with large neural nets...
Yes yes, we get it, Apple is not the preferred platform for playing games. This is not news, is not interesting, and will not change soon. What matters is that it's the preferred for people engaged in tasks other than playing games...
no, their performance claims have been spot on so far, just remember most games are not programmed well for Mac
the M1 Max already beat the RTX 3080 in Shadow of the Tomb Raider, so this new chip will beat the 3090 by a large margin in some games, and be much less in others
(just to be clear I meant it beats the 3080 mobile at a limited TDP like 100W or less, the 3090 is a lot faster, so the M1 Ultra being twice as fast as the M1 Max is necessary to equal the 3090)
I'm sorry but the M1 Max Struggles to even compete with a RTX 3060 mobile in games.
There has been plenty of tests done on this. The M1 Max GPU is only impressive in synthetic benchmarks. Doubling the cores is not going to help, it will still be slower than a 3080 mobile in actual workloads.
People are quick to try and point at drivers as the reason for low gaming performance is drivers or API calls. But this is just BS, these are not gaming focused GPUs. Back in the day AMD's GCN arch was capable of a good deal higher synthetic benchmarks than nvidia chips while performing around the same game wise.
The M1 is not impressive from a GPU standpoint, and couldn't even compete with something like a steamdeck on performance per watt even given the massive node advantage. The only area M1 is impressive is CPU and Hardware offloading. With the Level of integration, nearly everything being on the package allows for some pretty awesome performance per watt. Add on the massive Node Advantage, AMD nor Intel has any plans to go down that path.
Currently the money is on the Data Center and having access to many PCI-E Lanes for fast SSD Storage. Who knows when we'll get the Mac Pro M1 refresh, as the PCI-E Lanes just isn't there. Maybe Apple will address with with Future M2 Chips, but who knows. Seems like Mac Pro may again go away in favor of Mac Studio. And again back to the Thuderbolt addons.
Are you comparing native games or games running under Intel emulation? Benchmarks have been updated to run native. Everyone likes to compare Tomb Raider for some reason and that game is NOT native for Apple Silicon.
Ryan is literally one of *the* experts in the world on GPU benchmarking, and if he says that the performance claims are reasonable (assuming everything works as advertised), then I'd take his word for it.
Looking at benchmarks for the M1 Max, we already have a good idea how fast Apple's Mac GPUs are and since graphics is an embarrassingly parallel problem, we know it scales well as you increase the number of cores. Furthermore, Apple has a good track record with their performance claims.
I don't know why Apple does these comparisons, nor fanboys either. These Mac are workstations not gaming machines. These attempts at comparisons to something like a RTX GPU are pointless.
The comparison is to give people something to latch onto, to compare against something that (some segment of the population) are familiar with. That's all.
The point is not to give detailed comparison specs because, like you say, the actual target buyers do not care about games (or for that matter Cinebench); hence the vague graphs -- all that matters is to let potentials buyers know that "this is the approximate performance level; if that's what you're buying today, take a detailed look at what we have before writing that check".
There are applications that rely on GPU compute, and these benchmarks do give an indication of how performance is scaling. “Consumer-level” GPUs are frequently used for GPU compute tasks, too, so I think the comparison is relevant and fair.
And I completely agree with you that gaming is not at the center, Apple’s drivers are not optimized for gaming, so in many cases framerates are below what you’d expect from the hardware in many cases.
Why would it be optimized for a mobile GFXBench or "mobile architectures?" This is a workstation chip, and as such the GPU will be optimized for OpenGL libraries much like AMD FireGL and nVidia Quadro GPU's are.
People need to stop associating ARM and RISC architectures in general with mobile devices. Its important to understand a mobile GPU doesn't do anything differently than a traditional nVidia or AMD GPU, but everyone has a vastly different approach to architecture - mostly because of individual IP. Apple took care of that buying Imagination to protect themselves with a patent portfolio they can build on.
Because the SOC is based on Tile-Based deferred rendering.. it’s more efficient in some ways but worse in others. It’s also more power efficient, which is why it’s used in mobile hardware.
The M1 and it’s variants, as with other Apple SOCs, use TBDR.
gfxbench is optimized for TBDR. A TBDR based GPU would suffer a significant performance hit when you have post processing effects compared to GPUs that don’t use it.
Actually, TBDR based GPU would have an advantage in post processing because of their use of tile buffers to cache textures and color buffers for only the current tile, as opposed to immediate based GPUs that can't guarantee all the current data will remain resident on chip due to cache backing and cache contention as a result of multiple SMs competing for limited cache. RDNA2 mitigates this with a much larger cache. But a tile buffer is still more efficient energy wise - smaller buffer closer to the cores, no need to worry about cache conflict.
Nvidia has been doing this since Maxwell. That's how they managed to stay competitive with Fiji (i.e. "Fury"), which significantly out-classed it on raw specs.
AMD has been playing catch-up, here. They added a hardware feature called DSBR (Draw Stream Binning Rasterizer) to facilitate TBDR in Vega, although it remained unused until late in Vega's product cycle. When finally enabled, I think I read it was good for maybe a 10% performance boost.
With Apple, R stands for rendering. Their whole pipeline is tile-based. You can watch WWDC videos on how their GPUs work, they've made several in 2020 and 2021.
Without digging into that, it seems a bit silly to say "rendering" is deferred, because the entire process is collectively "rendering". Perhaps they simply wanted to emphasize that more than rasterization was deferred.
> Their whole pipeline is tile-based.
It's basically what Imagination Tech used to do. They were tile-based since the early days.
Definitely super impressive by Apple. I'll wait for benchmarks but what Apple is doing will hopefully push the rest of the industry to have better laptops and power efficient designs!
Apple likes to fudge their numbers and make it seem like magic, but its hard work by engineers and process advantages that Apple holds.
Welp, I hadn't guessed we'd see the first MCM GPU from Apple, but apparently here we are. It will be interesting to see how it actually performs in practice, but it seems that most review houses have been having trouble finding like-for-like GPU workloads to compare between the Apple ecosystem and anything else.
You assume that Apple sticks to the same product cycles as when they were bogged down by the lackluster intel performance increases. That Apple can keep up a steady 12-18 month product cycle where it matters and they have full control should be obvious from all their i(Pad)OS based devices. So, I’d bet, your assumptions are not only wron, but also a psychological defense mechanism, because the prospect of Apple dominating desktop performance is a concept you haven’t made peace with, yet. You will, in time, have little choice…
For an otherwise reasonable post, this is an unnecessary hit below the belt. Let's try and stick to facts and logic, while accepting that reasonable people can disagree and each is entitled to their own opinion.
Whether or not it truly beats RTX 3090 is kinda besides the point IMO. That is probably not as important to Apple as having ALL of the M1 Ultra machines with the exact same very high GPU performance, at a much lower power budget. And don't forget that this is effectively a small form factor design. The top end Mac Pro hasn't even been announced yet. I am guessing that will be announced in June at WWDC.
The Mac Studio is a iittle pocket monster that I'm sure will be super popular with creative professionals. The $4k version has both a more powerful CPU and GPU than a $15,000+ Xeon Mac Pro.
But it doesn't have any expandability outside all those (admittedly super fast) thunderbolt ports and doesn't have user upgradable RAM or ECC RAM. What is even more interesting will be the next Mac Pro, which is rumored to have dual sockets for two M1 Ultras, along with expansion slots and a big chassis. Will it support expandable ECC RAM? It's possible they might provide expansion through CPU swaps. Start with a "low end" single 64 Gb Ultra, later pair it with a 128Gb high end Ultra when you need more power, and eventually when it's cheap in a few years swap out the low end Ultra for another high end one.
Never going to happen. Apple will not engineer an LGA socket for their uber niche market, this thing won't even be bought by masses. Look at the price tag, $1500 base for pathetic M1 Max. And 4K for the real upgrade and to beat a 3090 you will need to add 2K more to make it 6K and then even more for more storage and memory.
Mac Pro will not come now, I think it might be axed or maybe Apple just for bragging rights will do one for M2 series. As you mentioned this cannot be a workstation class machine. Its too niche and too expensive plus everything is soldered.
It seems to me that they need ‘two different chips’: the original M1 Max and a ‘mirror copy’… so the can connect the side wires in the correct order.
If we have one chip with —say— the connector over the right side and the first wire is at the top… when we rotate the other chip so the connector is on the left… the first wire would be in the bottom…
…except that the ‘bridge’ inverts the 10.000 wires.
About half of the pads are for transmitting data and the other half are for receiving data. You don't want to connect the same pads together otherwise you would be connecting transmit pads to other transmit pads and receive pads to other receive pads.
Usually, the chips alternate transmit/receive pads (e.g. pin 1 transmit pad 2 receive.... pin 1023 transmit, pin 1024 receive) and then when the chip is flipped around, one chip's transmit pads line up with the other chip's receive pads (e.g. pin 1 transmit -> pin 1024 receive, pin 1023 transmit -> pin 2 receive.)
Everything falls apart for singing praises for this garbage design which Apple excels at best - BGA solder.
This processor is having $1500 price tag for a basic M1 Max SOC and with $4000 cash you get the Ultra upgrade, but the DRAM is unified so the max memory for 4K is 64GB, and that will eat into the GPU memory, okay it's huge and very fast. Still it's nothing ground shattering since how Apple's M1 Max and Pro got slaughtered in dGPU tests last time at Anandtech, I remember how they lose to a crappy mobile GPU which is 3080L / 2080 Desktop class on Aztec Ruins and in SOTTR it crawls, at 3060L lol. Awful garbage.
Now we have this omega marketed halo product which commands such a high price tag with similar expectations and for the dGPU performance it will again be Apple OS based vs the GPU rendering power, and since ARM processors use specialized blocks and this one has more blocks for decode and encode it will be faster as Apple claims however the advantage is very thinly spread due to it's niche nature.
Now since it's already mentioned how Apple SOC is massive in area and the Transistor count, and the most funny part is look at the CPU and GPU performance 12900K and 3090, now they say it's almost at that level for the most part and GPU is where they beat and ofc it's all similar performance only advantage where they beat is Power consumption. So ultimately pay $6000 USD to have a BGA soldered fancy looking box which will have exactly similar performance to a latest PC LMAO. And ? We are supposed to take this rubbish seriously ? Yeah I guess.
Meanwhile Z690 or any Mainstream mobo has PCIE expansion slots for vitually anything, and a lot of I/O instead of the BS Thunderbolt ports which again needs adapters to have multiple ports extra money needed. And the worst part is even the Storage is soldered. Like pay 4K + extra 1-2K for 4TB, for that amount I could buy Intel Optane P5800X which will shred this pile of junk to ash and outlive this silicon, OR I could simply buy FireCuda NVMe SSD which is made of Micron's latest flash 196L and gets me insane TBW endurance and performance, ofc I can keep it cool because I use a Z690 motherboard.
As for Workstation class, this product is a giant failure. You cannot run any sort of VMWare Hypervisor home server class software OS, there's simply no way this junk will support that, on top you have to spend a ton of cash. Why not simply get an office decommissioned XEON processor and a good PowerEdge / ProLiant HPE chassis and add HBA SAS expander and blow the roof out of the RAID ? Or say for eg buy the new upcoming Threadripper series which is a HEDT machine and will churn everything you throw at it and will live super long long time and on top you can get a TON of DRAM memory plus the massive I/O bandwidth ?
Yep, I will forget everything and simply simp for this junk company and their bragging of TSMC first hand product access. Wake us up when AMD and Intel and Nvidia get TSMC 5N lithography access and let's see the M1 Ultra get slaughtered.
I don't think it's entirely uncalled for. While Apple has some technology that I'd normally be quite excited about, the level of excitement is quickly checked by it being... Apple. A $4,000 system that is completely nonupgradable, nonmodular and unrepairable, with even the SSD being soldered? The only way you can get behind that in good conscience is it were at least twice as fast as the next best competitor, but it really is not. And with all their grandstanding at every turn about their commitment to "the environment" while still maintaining absolute resistance to any form of independent repair, I'm not sure calling them a "junk company" is at all unfair.
It's a great big shame, really. There's a lot of things in the M1 series that I'd like for myself, but it only being available in Apple products just kills it for me.
I like where a lot of the M1 is going. I'm curious how much is better design, how much is just having an enormous transistor budget, how much is better manufacturing, and how much is false marketing.
For example, comparisons to nvidia products made on samsung's process aren't fair from a power consumption point of view at all. The samsung process is way worse and the 3090 wasn't tuned for power consumption, it's a massive card with a huge heatsink and cooler.
The massive memory bandwidth and faster integrated graphics is something I've often wondered why AMD doesn't pursue in a laptop/desktop form (like a ps4/ps5) as clearly they know how. It's not clear to me why Apple is doing it first but good for them. Personally I would probably be willing to trade off upgradable ram for better performance if they don't price gauge you for it. For SSD's I can't see any valid reason to make them non-upgradeable as PCIE4/5 isn't holding them way back. To date I've never seen a vendor that doesn't price gauge on RAM and SSD upgrades so I'm not hopeful in this regard.
I also see some misreporting, eg there's an article on Toms claiming a 3990x threadripper is the same on geekbench but if you look at the linux (not windows) results for the 3990x it's 40% faster. There are also claims of 3090ish performance but then independent testing shows way lower results most of the time. And something like ethereum mining is dreadfully slow on the m1 variants by all accounts, etc.
At the end of the day I still think it's probably the best laptop chip you can get but I hope we see some more competition.
> massive memory bandwidth and faster integrated graphics is something I've > often wondered why AMD doesn't pursue in a laptop/desktop form
$$$
The consoles and this Apple Studio Pro are vertically-integrated, which means they're not stacking margins on top of margins. Furthermore, MS and Sony lose money on the console hardware and make it up in software royalties and subscription services.
If AMD makes a monster APU, it's going to be expensive to the point where it'd be a niche product and most in that price range would prefer to get just a CPU and dGPU.
Worse, it would limit you to non-upgradable memory, since you can't get the needed bandwidth with DIMMs.
> clearly they know how.
Right. They *do* know how, and haven't seen fit to do it for the past 10 years since they started making console APUs. That's your clearest sign that it hasn't been viable for the mainstream market thus far, because they've been keen to play up their GPU advantage over that entire time.
SSD is replaceable. Macs last forever. It's got a ton of super high speed thunderbolt ports for all the external storage and devices you want. And its hands down the fastest video and graphics production system you can get under $5k.
>SSD is replaceable Only with original Apple parts (since it's not a standard NVMe drive), which not only means that you cannot upgrade it with third-party options, but also that you can't take the SSD out of an otherwise potentially faulty system and get the contents out of it in another system. >It's got a ton of super high speed thunderbolt ports for all the external storage and devices you want. What if I don't want external storage, but just larger internal storage? Or RAIDed internal storage? Or more memory? Or a 32-core GPU with only a 10-core CPU? Or to upgrade my CPU (or the WiFi adapter) three years down the line? And so on and so forth. >And its hands down the fastest video and graphics production system you can get under $5k. Only because of the media engines, and there's absolutely no reason they couldn't be on something like a PCIe card. >Macs last forever At this point I hope you're trolling.
“Repairability” is the biggest cause for necessitating repairs. It is much better for the environment to toss one and recycle one in e.g. every 10’000 units than to toss/recycle zero units, but have to repair one in 1’000. Apple has the volume to make these statistics, and they optimize globally not based on the emotions of a single affected consumer. Reliably of glued-down, soldered-in-place parts is much higher than that of removable, socketed parts. While the latter can be repaired, stock kereping, packaging, shipping of repair parts has a significant environmental footprint, too; and multiplied by the higher failure rate, the environment takes a beating compared to the non-repairable approach.
There are valid points on either side. Neither approach is going to be right for everyone, and some people just value the expandability of a machine that's designed to be upgradable.
I have a workstation with a NVMe SSD in a PCIe slot. It has no M.2 slots, but I have enough PCIe lanes that I could add a PCIe card when they came along. I like being able to do that sort of thing. I also upgraded the GPU in this machine, which was a huge benefit, though I also had to swap around PSUs when I did. I was also glad to have that option. Another mid-life upgrade I typically do is to double a machine's RAM. Sometimes, I've even done 2 RAM upgrades, doubling capacity each time.
At work, I've upgraded machines' CPUs, RAM, and GPUs to fulfill project needs.
So, as much as you might favor a certain approach, you really can't say it's right for everyone and every purpose.
>Reliably of glued-down, soldered-in-place parts is much higher than that of removable, socketed parts. Do you have a citation for that claim? It seems immediately false to my experience. Extremely rarely have I had such components as socketed CPUs, GPUs, RAM or SSDs fail, and even less so in a way that is traceable to them being socketed instead of soldered. Intrinsic failures appear to be much more common with such things as capacitors or other simple components that are merely part of something like a motherboard or PSU, and in reality failures are much more commonly caused by some sort of user error or other external factor (liquid damage, dropping things, power surges, and so on), which would affect a monolithic system just as much as a modular system. >It is much better for the environment to toss one and recycle one in e.g. every 10’000 units than to toss/recycle zero units, but have to repair one in 1’000. To begin with, I don't believe that's true, unless you have some citation for that as well. It seems obvious to the naked eye that a proper and limited repair has a significantly more limited impact. Changing a bad cap or cleaning up liquid damage is literally nothing compared to recycling and assembling a whole system. I also don't believe that modular systems fail more than an order of magnitude more often than monolithic systems. Nothing in my experience tells me that would be the case. See above, I guess. >Apple has the volume to make these statistics, and they optimize globally not based on the emotions of a single affected consumer. Is your argument actually just "trust Apple, they're just such good guys, they would never do anything to screw over the customers in order to improve their bottom line, it just looks that way to you because they are optimizing for global variables"?
Not that I'm raving about the Studio but you do realize with a small form factor NUC design you would slave a Thunderbolt Hub off the Studio, then slave off a NAS and be done with it.
I agree the Zen 4 will turn heads on the 5N process this Fall.
Those are certainly all words. " Like pay 4K + extra 1-2K for 4TB" Doesn't seem entirely unreasonable. Go to Lenovo's website and configure a P620 workstation and see how you can pay $1K for 2TB of NVME. $1K per TB is how much I expect to pay for enterprise storage solutions anyways.
"you cannot run any sort of VMWare Hypervisor home server class software OS". Lol who cares? We run our hypervisors on our datacenters on proper servers, not workstations that live on employee's desks.
"Why not simply get an office decommissioned XEON processor and a good PowerEdge / ProLiant HPE chassis and add HBA SAS expander and blow the roof out of the RAID" Buy used, slower hardware instead because it's cheaper?
"Or say for eg buy the new upcoming Threadripper series which is a HEDT machine and will churn everything you throw at it and will live super long long time and on top you can get a TON of DRAM memory plus the massive I/O bandwidth ?" I guess that's always an option, but expect to pay a lot. Lenovo still charges $4500 to upgrade to a W3975WX, and you're still left with significantly less VRAM no matter what discreet card you go with, and consuming multiple times more power, from a much larger device.
Not sure why you're hating so much - M1 Ultra and Mac Studio are very price competitive and offer incredible perf/watt, and it's not even the highest tier M1 series SoC that's going to be released.
I think everyone sane. And their mother, grandma, plus their dog, want something repairable and upgradeable. Not some premium junk-that-when-break-go-straight-to-bin like these M1 what-what-ever. Soldered RAM still ok (although you must pay some hundreds bucks for 2xRAM size, lol). But soldered SSD?! Damn Apple and their fanboy that said it's OK configuring something like that.
Same with x86 soldered and unrepairable junk. Like most of x86 chromebook and MBA-wannable-ultrabook.
Macs last forever and depreciate slowly. When I want more performance than my M1 air can offer, I'll sell it (for still a very good price) on Ebay and buy new.
This is a huge step forward for chiplet technology. Another step forward would be edge-to-edge interconnects on top of the 2.5D interconnects. Those interconnects are even faster, and can support multiple signal types, such as analog/digital/optical/power all on the same edge. An Indiana company is developing that: www.indianaic.com.
With 114B transistors you can build 6900XT+RTX3090+3*5950x Also clocked low and using 5nm TSMC. it would be ridiculous not to be more effective, but that ladies and gentlemen is how to laste transistors on closed system in time when transistors are not enough for everyone
Funny how many people are saying "unbelievable." That's what people said last year, when a juiced up mobile chip kicked the behinds of every desktop chip except for the highest end ones. When it comes to reality, your belief is not required.
The heatsink on the Max is aluminum. The heatsink on the Ultra is copper.
Copper weighs more than aluminum, but is better at moving heat, which is why coolers often have copper heat pipes even when they have aluminum radiators.
Teraflops numbers can’t be compared across different architectures. It is a simple figure that are calculated by timing frequency and the numbers of ALUs together. How will are these ALUs been used to be turn into real world performance was not reflected by these figures.
Before the introduction of NVIDIA’s Ampere architecture. NVIDIA used to have a slight lead with relative performance per TFLOPs. Their 1000s series GPUs can outperform AMD competitor at the time despite having lower TFLOPs. However, with introduction of Ampere comes big architecture changes. NVIDIA decided to double the amounts of floating point ALUs by having the newly added ALUs share the pipeline with integer ALUs. This mean half of floating point ALUs on Ampere GPU have to choose to compute either floating point OR integer calculation, but not both simultaneously. This mean not all floating point ALUs can be utilized at any given time if integer operation was required. Also other bottleneck might exist in the architecture to properly feed twice the amount of floating points units. Regardless, NVIDIA didn’t achieved doubling the performance of traditional raster or shader performance with this modification, this mean now NVIDIA’s TFLOPs number are highly inflated compared to relative performance.
AMD’s 6900 XT has 23.04 TFLOPs, but it can trade blow on raster performance with RTX 3090 despite the later having 35.58 TFLOPs.
It looks like someone doesn’t understand the difference between theoretical and actual performance. Integrated systems come much closer to theoretical performance than discrete GPUs do.
The first multi die GPU was released in 2021 by AMD. Are you saying the first prosumer GPU to use the tech? CDNA2 is in multiple AMD GPUs currently used in data centers.
This is more due to software than hardware though. The interconnects for infinite fabric with the MI200 allow for up to 800 GB/sec between the two chips. AMD chose to have them present as two GPUs; Apple has chosen to do one. We have no benchmarks or data to show if apple's solution will be effective or not.
It compares fine when it's just a GPU. That 2.5TB/sec also accounts for a CPU, Neural processor, and more, the GPU physically can't take over all of that bandwidth.
I would imagine that most of that bandwidth would be used to access system level cache on the other die. Since the system level cache is accessible by CPU, GPU, Neural engine, etc, and communicates with the RAM controllers (both chips have RAM channels), I would think most of that bandwidth would be available to the GPU or CPU.
> That 2.5TB/sec also accounts for a CPU, Neural processor, and more, the GPU physically can't take over all of that bandwidth.
It's extremely unlikely to utilize all of those units to the max, all at once. That's a corner-case benchmarking scenario, not something that happens in the real world. In the real world, whatever you're doing might max out one thing or the other, but the app is going to be bottlenecked by aspect of the hardware, rather than have a perfect balance across the entire SoC's engines (if it even *uses* them all!).
What this means is that Apple basically just needs to design the SoC to support the most demanding aspect: CPU, GPU, or Neural (plus a little headroom, if they can manage it). Not all three, simultaneously. And if you did manage to max all 3, then I'm sure the SoC would do some significant clock-throttling.
> infinite fabric with the MI200 allow for up to 800 GB/sec between the two chips.
Incorrect. There are 4 links between the dies of MI200. Each link can do 50 GB/s per direction. So, the aggregate is only 400 GB/sec. That's only a quarter of each die's memory bandwidth. So, if GPU cores are naively pulling data out of the companion's memory, that interconnect would be a crazy bottleneck.
That's probably why AMD *doesn't* treat them as a single GPU. They're merely presented as 2 GPUs that happen to share a package. MI200 seems primarily designed for use in systems that have more than one. And if you have a compute workload that scales to more than one MI200, then it probably doesn't hurt to treat the GCDs inside as separate processors.
To be fair, AMD’s MI 200 is a lot more primitive when compared to this chips though. MI 200 only had up to 100 GB/s of bi-directional bandwidth between both GPUs which can be achieved with NVIDIA’s NVLINK without any multi chip packaging. This chips had 2.5TB/s which is very close to 3TB/s found by NVIDIA’s researcher to be the optimal inter die link bandwidth in their MCM research paper published in 2017, so Apple’s claimed that this multi die GPU will act as one monolithic GPUs may have some based in reality. If Apple actually pull that off, they will be the first one to crack the puzzle.
No, that's double-counting. Each of the two chips devotes 4 of its 8 links to its peer, but you only get 4 links' worth of bandwidth because they're connected to each other.
For a system with only one MI200, you could get 800 GB/sec between them, by connecting up the external links between the two GPUs, as well.
> ‘Apple is understandably excited about their accomplishment’ > ‘its’, ‘the company’, or ‘the associated staff at Apple’. Each of those is appropriate but not ‘their’. > Corporations are not people, my friend.
That's a strange claim, especially considering your nick... See here: https://www.grammarphobia.com/blog/2018/01/it-or-t... In AE, companies are singular in *formal language*, but plural in more informal contexts. In BE, however, companies are much more likely referred to in plural also in formal texts.
The Fiji chip came after nearly 4 years of 28nm products being on the market and all the process improvements achieved over that time. The N5 process is barely over a year old so the defect rate is probably still high enough for a 432mm2 chip to be considered as on the verge of manufacturability.
Yup. The 119 mm2 M1 die is tiny in comparison, but defects on N5 did start out high enough that Apple sells devices with a disabled GPU core at a discount instead of tossing chips with a defect in one of the GPU cores out.
If you look at the Ultra, they are offering an option with 48 GPU cores enabled and another with all 64 GPU cores enabled, so wafer defects are still a thing on the two 420 mm2 M1 Max dies that are combined into an Ultra.
Probably right, except Apple probably doesn't want to make a laptop that big & chunky (even the current 32 GPU M1 Max will throttle at extreme loads in a laptop the size of the current MacBook Pros).
I expect the new Mac Pro to have M2 Max chips with Ultrafusion fabric connector on TWO or THREE sides. This allows FOUR to EIGHT M2 Max chips to be connected into one monolithic processor! The GPU of the M2 will bring Ray Tracing to the Mac, finally. And the M2 Ultra will have between 128 to 256 GPU cores! This would be a MONSTER chip as befits a Mac Pro.
It's good, but how can I easily use it like an i5 + Radeon I have? At nothing ! 1.PC Gaming = NO 2.Audio: HiRes Player + Dac + Streamer + Amp = NO 3.Working: Autocad + SolidWorks etc etc ETC = NO 4.Working at Home: Team V., Any Desk, VPN, Remote D. etc = NO 5. Own Datacenter server accessible externally from anywhere with Remote D. = NO 6. Navigation anywhere without restrictions = NO 7. A very good protection available to be managed: Bitdef. Total Security etc etc = NO 8. 100% compatibility with the whole world: word, excel = NO 9. All Social Media very easily on my desktop: Whatapp, Instagram etc = NO 10. All movies and music very easily on my desktop: Netflix, Spotify = NO ... and of course a simple mouse with 2 buttons and scroll on my desk = NO atehnic magic mouse
Damn I didn't realise that you can't use a 2 button mouse with scroll wheel on a Mac! Oh, wait, you can. I could go on but I think I can sum your post up with just one more word....
I used a Logitech wireless mouse on an Air - a bad experience, then an ergonomic Microsoft wired mouse - another bad experience, then an Apple Mouse Magic - the worst experience (too small for my hand, inaccurate in sustained work ) ... plus a non-IPS screen - something very unpleasant overall ... I gave up. Otherwise: very fast, AL case super ok, excellent os (superb) ... no comments.
So you bought the cheapest possible Apple laptop, then complained it’s not IPS? But at the time that fact would have been obvious to anyone who bothered to check what they were purchasing.
Would you buy a PS5 then complain it doesn’t run Word, despite the maker not making that clear anywhere on their website?
In any case all Apple laptops now have IPS screens even the cheap models.
Logitech mice work fine with macs & so do all USB mice & keyboards. Maybe you struggled with the different mouse acceleration curves between Mac and Windows - that can be a big thing for some people (it is for me). Apple mouses have always been shitty, I don’t touch them. I’ll give you that one.
But that has no bearing on the list you gave. I only know half of the use-cases you mentioned but the half I do know about are bullshit.
Eg 7: ‘navigate anywhere without restrictions’ do you mean navigate around the OS? Yes you can turn off the restrictions any time you want or sudo / root your way in or even turn off sudo / account password entirely. Complaining about that ‘restriction’ is like complaining that Windows UAC illegally prevents you from doing whatever you want. It’s an ‘are you sure you know what you’re doing?’ step, but done rather better than UAC / Registry.
Your list is somewhere between satire and "has almost no idea about Macs at ALL but enjoys writing about them in comment sections anyway". I'm going to lean toward the latter.
🤣 Of course you can, just about everything! Just because someone doesn’t ship with the box or (temporarily still) runs in Rosetta2 doesn’t mean you can’t use or do it.
I don't see multi-GPU as much of a hurdle for Apple. In what workloads are users going to be sensitive to GPU latency? Gaming is an area where people will freak out about less than perfect frametime consistency. Who is going to game on an M1 Ultra? AMD and NVIDIA have their work cut out for them to produce good multi GPU gaming chips.
I agree. Gaming has just NEVER been a priority for Apple and I don't foresee that changing even with their recent moves a la Apple Arcade. The M1 Ultra may well be an awesome chip for gaming but I don't think we'll ever know.
Gaming on iOS is huge from a market share standpoint.
Going forward, games on ARM based Macs will run on the exact same CPU and GPU cores using the same graphics framework that iOS games use. This means that targeting the Mac is no longer a niche with little market share. The same game code targets iOS and ARM Macs.
This worked well for XBox when it was breaking into a gaming console marketplace that Sony dominated, since XBox leveraged the same Direct 3D APIs, CPU cores and Nvidia GPUs that Windows games already worked with.
XBox had a tiny console marketshare compared to Playstation, but if you were already targeting Windows, it wasn't much more work to target XBox.
The same will be true of studios targeting iOS devices. Targeting ARM Macs is not much more work.
You’re right, iOS could well be (by numbers of devices) the largest gaming market in the world. How’s that affected Mac gaming? I don’t think by much, certainly not that I’ve noticed. Don’t get me wrong, I’d LOVE to see Mac gaming take off and was as excited as anyone about the possibilities when the M1 was announced. It just hasn’t happened. I hold out some hope for an Apple Silicon games console but let’s be honest, that’s probably not going to happen is it?
It hasn't effected Mac Gaming, because until now, coding Macs and coding iOS devices have been two different things.
That is no longer true. ARM based Macs can play iOS games unchanged with zero effort on the part of a developer.
However, it makes sense to put in the minimal effort needed to at least add support for a game controllers (that support will work on iOS devices too) and support for Mac style menus.
So, sorry for your favorite narrative, but ARM Macs are going to be supported in a way older Macs never have been.
I’m a Mac user and have been for over 25 years. Our household is a complete Apple household from the 27 inch iMac in the office, the M1 MBP I take to work, the two AppleTV’s, 12.9 inch iPad Pro, in fact multiple iPads Pro, multiple iPhones, HomePod speaker, original Apple HiFi speaker, iPod Classic (somewhere), and the many old and not used anymore Macs I have lying around.
This isn’t a narrative it’s a fact, despite everything Apple has done developer’s just aren’t developing games for the Mac in the numbers that You, I, or anyone interested in gaming would like them to.
I agree that the AS Macs are likely to be better supported by Devs but so far they haven’t been the panacea we expected.
Yeah, it's unfortunate that Apple just has no real interest in the gaming market. Sadly, they think crappy indie games on Apple Arcade checks that box. I suppose it does for the Candy Crush crowd, but not for actual gamers. I can't help but look at the Apple TV as such a missed opportunity for them.
As I contemplate purchasing a Mac Studio, I have to wonder whether the M1 Max is sufficient enough. For video / photo editing, the M1 Max is more than enough. If only there were some really high end games where the Ultra would make a real difference...
I am reminded of the people who stuck to the narrative that AMD couldn't possibly compete with Intel back when they hired Jim Keller and put him back in charge.
We are now moving into an era where Mac gaming is changing from a not terribly profitable niche to Mac gaming piggybacking off of one of the most profitable areas in gaming.
Cheap ports to a ARM Macs changes the economics of game development.
I’ve been around long enough now that I know anything is possible. From the original days when nobody believed AMD could compete with Intel - and then came Athlon and 3dNow, then when AMD fell behind again that was supposedly their end, they were doomed. Same with Apple, I bought into the Apple ecosystem around the start of OSX (my first laptop came pre installed with both OS9 and OSX) when everyone was saying it was all over for Apple, and never bought another PC since.
But I’ve also seen Apple promise time and again that “this time we’re serious about games” and rolling out the likes of Jon Carmack to prove it only to see Max games withering as always.
I agree that this time they have a real chance at it, but to do so they need to be pushing a box like an AppleTV Pro with the latest Apple Silicon in it and a crap load of onboard storage - and keep it updated, every single year.
But they haven’t, and I worry that game devs have been burnt once too often to commit serious resources to developing AAA games on Mac.
I don't know about "zero" effort. Many iOS games would lose a lot going from a small touchscreen to a large non-touch keyboard/mouse interface. But more to the point, what iOS game is going to benefit from the power of an M1 Ultra? What person would buy an M1 to take their iOS gaming to the next level? Now ask those same questions about NVIDIA and AMD GPUs. Gaming is the primary driver for PC GPU innovation and barely a thought for Apple M1 GPU development. Game frametime consistency and latency are the main reasons we don't have multi-GPU gaming today; it's one of the reasons SLI died. The intended workloads for M1 Ultra don't have those concerns.
>Many iOS games would lose a lot going from a small touchscreen to a large non-touch keyboard/mouse interface.
Yes. Just like PC Games based on a mouse and keyboard didn't work without some effort when moved to an XBox. You had to add support for a gaming controller.
Adding support for gaming controllers on an iOS game means those gaming controllers (XBox or Playstation) will also be supported on an ARM Mac.
As I said, there is not zero effort required to make a good Mac game, but it is definitely minimal effort.
Not to mention gaming engine makers who can't support iOS without also supporting the Mac now.
I'm curious about their multi-gpu for gaming as well - I suspect it hasn't solved the hard problems. It's not like it's a unique or new idea, I had Voodoo2 SLI like everybody else back in the day. But that technique didn't scale to modern games. ATI had a rage fury maxx that was a dual chip AFR render (which added latency) and there was the nvidia SLI of the past decade, and various other AMD implementations which were often game dependent. Both NVidia and AMD have said it's not as simple as putting two chips on the same card. If it were a 3070 would probably be 2 3060 chips and a 3080 would be 3 or whatever.
#applesempretroppoavanti Apple manages to overcome each time, this sector is very complicated and full of pitfalls but thanks to the determination of the Company, they have also managed to become successful in the world of chips and thanks to this new way of connecting the 2 dies they have really raised the rod. the Apple Silicon product was a revolution and also the new ultrafusion system will be a reference for the entire industry.
Great performance, I'll give them that. But running 128GB of RAM without ECC is scary. Maybe the graphic/video people don't really care about a bitflip once in awhile.
on-die ECC is, to be nasty, a means to increase chip yields, not to increase reliability. That said, the key thing of ECC is that the OS knows it has verified data, and for that ECC has to operate during RW ops, not during refresh, and the RAM needs to be able to signal failures to the OS, not silently fail. Especially when running modern file systems this is key, because it’s easy to corrupt large swaths of storage by writing corrupted fs metadata
> on-die ECC is, to be nasty, a means to increase chip yields
Overall, yes. I get that. However, it's conceivable that Apple might've contracted with a DRAM supplier to allocate more on-die ECC bits so they can have a net-increase in reliability, no?
> the RAM needs to be able to signal failures to the OS, not silently fail.
Depends. If you have replaceable RAM, then it's valuable to know that you have a DIMM that's beginning to fail.
The other reason is so the OS can know a certain address range has a defective memory cell and it can reserve that page so it's taken out of circulation. However, this *could* also be handled on-die, where the DRAM itself could remap bad rows. It'd be far better for the OS to do it, of course, so that the DIMM didn't have to implement physical address translation.
I can’t see how AnandTech can claim this is the ‘last member of the M1 family’. The Apple Silicon transition is not yet complete. We still have the Mac Pro to go.
That could be the M2 or it could be yet another version of the M1.
Apple very explicitly said in the presentation that the Ultra was the last member of the M1 family of chips. So AnandTech aren't claiming it, that is what Apple said.
Thanks for the clarification. That phrase 'We're adding one last chip to the M1 family' wasn't in the AnandTech liveblog, but Ars Technica reported it in their liveblog.
That's interesting, because speculation is that the Mac Pro will have 2x M1 Ultras glued together - that lines up with all the previous rumours about the new Mac Pro's upcoming capabilities. But every time Apple have stuck two chips together, they've given it a new name. M1 Ultra = 2x M1 Max = 4x M1 Pro.
So either this is NOT the final (and apple didn't use the word 'final') M1 chip, OR it will be called the M2 (unlikely) OR it will be 2x M1 Ultra chips in separate packages, not connected together.
Using 2 separate packages seems un-Apple like, and introduces unnecessary complexity / issues. So my money is on the first option - that Apple will introduce yet another M1 variant for the Mac Pro. "We said in March that we were adding one last chip, the M1 Ultra. Well, guess what, we've made a new one! Witness the M1 <Plaid>!"
Historical evidence: (1) Apple constantly one-upping themselves each year. (2) M1 Max being inferior (!) to the later release of the M1 Ultra.
Connecting two M1 Ultra is no more complex than connecting two GPUs in the current Mac Pro. At the worst they use PCIe5 x16 for 128GB/s between chips and at best they use UCIe and if they can keep the two chips 2mm apart they can connect them at 1.2TB/s
It seems far more likely to me, as already discussed here, that they’ll wait for the M2 for the Pro and that the M2 will have extra interconnects to allow them to combine more than 2 dies at a time.
That might be a struggle within their original migration target, but it is entirely possible that they might pre announce the new Pro sometime before it is available, giving them a bit more time.
that's very unlikely because Apple will never give up scale economies for M2. they obviously went this path with the M1. they started with iphone and by the time the A13 (or whatever the number was) became M1 they had produced millions and brought production costs down. then they came up with M1pro/max when those became cheap to make. and so on.
they will introduce the M2 on Air/mini, then MBP 14/16, then iMac, then Studio, then Pro, which I'm guessing follows the sales numbers of the respective devices (save for the mini, I guess). there would be a huge cost in starting M2 sales with the Pro.
It's also possible that the chips in the Mac Pro line will be considered a different chip family altogether.
One that focuses on throughput without caring about power draw or production costs.
I wouldn't be surprised to see massive amounts of HBM,for example, that would not be cost effective in a consumer product like an iPad, but is already in use in high end products from others.
Is it possible that synthetic benchmarks are currently capable of accurately Apple SOC chip architecture? When you add encoders/decoders, unified memory, neural engines I think they underestimate real workloads.
Either Apple does not know how to label a graph properly, or I think there are some major misinterpretations going on here.
The GPU performance vs. power graph.
Y-Axis = "Relative Performance." Not "Performance." It says "Relative Performance."
Relative to what? It's clearly not relative to the 3090 setup (otherwise the 3090 results don't make sense, it would have to be a straight horizontal line). From how I read it, "Relative Performance" is "Performance / Watt." Let's call it P/W. You can think of P as Performance Points, kind of like a 3D mark score or something.
Let's look at the maximum performance on the graph. The top of the M1 Ultra looks to hit about 108W (just a little below the midpoint between 100 and 120 on the X-Axis).
So at 108W, the M1 Ultra hits ~180 on Relative Performance (P/W). So 180P/W x 108W = 19,440 Points.
The highest point of the 3090 setup is about 315W. The way the graph reads, I'd peg this around 170P/W, but lets use the 90% figure they also used on the CPU graph, and give the Apple chip even more of a lead, (180P/W x .90) and we get 162P/W for the 3090 setup.
So at 315W, the 3090 set up is ~162 on Relative Performance (P/W). So 162P/W x 315W = 51,030 Points. (It would be even more if I used 170P/W)
So the M1 Ultra is still very impressive, scoring about 38% of the 3090 setup (19,440 / 51,030), at only 34% of the power (108W / 315W).
Of course, if they actually provided some sort of units or benchmarks for the Y-Axis, they could clarify this whole mess of a graph. If it really is that fast, label is properly!
Don't these numbers seem much more reasonable? And it seems to fit better with how previous chips compare with low TDP versions of the graphics cards (mobile 3080's for example).
Thoughts? I feel like Apple is intentionally misleading us with that "200W less power" notation.
"Relative performance" does not mean performance per watt. "Relative performance" just means that the ratios between the two chips are accurate but it does not have a specific unit. This is very common nomenclature in science when comparing two processes. It is probably based on a summary of their internal benchmarking process scaled to make the numbers pretty.
They're lovely chips but I really really wouldn't bother buying a Mac Studio for gaming use. You need at least $4000 to get an M1 Ultra, it won't run Windows well (yet) and there is no guarantee that any Mac games will be properly coded to make use of the M1 Ultra power.
Better to, if you can, pay for gigabit internet service (£25 a month for me in the UK with 2 ms ping) plus a subs to GeForce Now ($20/month or $200 per year) for 3080 tier gaming on your Mac / PC
I could actually see Apple taking someone else's Neoverse V1-based server CPU, for use in the new Mac Pro. Maybe they'll put an M1 Ultra on a PCIe card and use it as a dGPU?
A YouTuber did some rather sound research, including paying for paywalled articles from TSMC which outline the interposer tech used in the Ultra, and an additional version of the interposer that supports four-way linkage, plus a data backbone that will be used to connect to slower large scale RAM banks. Likely based on M2, so 48 cores (8 extra efficiency cores), 160 GPU cores, and up to 256GB unified memory with access to a further 1TB of more classical RAM. Also stated that the same backbone will be able to access additional GPU cards, and suggested that Apple may sell their own GPU PCIe cards like they did with afterburner.
> including paying for paywalled articles from TSMC which outline ... > an additional version of the interposer
If, at this point, your Spidey senses don't tingle, they must be broken. There's no way TSMC is going to dish sensitive details about their Apple-specific solutions. And there should be no doubt that this interposer is *very* Apple-specific. But nobody is going to fact check it, especially if the sources are supposedly pay-walled.
Unless a leaker has a strong track-record, I pay them no mind. It's too easy to dress up mere speculation and the incentives are tilted strongly in favor of the sensational, with almost no negative consequences for being wrong.
" A YouTuber did some rather sound research " which one, post it here so we can all go and see it. sounds like pure BS to me, which is what it looks like Mode also implied.
quote "Apple’s die shot of the M1 Max was a bit weird initially in that we weren’t sure if it actually represents physical reality – especially on the bottom part of the chip we had noted that there appears to be a doubled up NPU – something Apple doesn’t officially disclose"
It looks like the old die shots were deliberately misleading - it's easy to compare the old and new images. Where the "2nd NPU" was appears to be where the 'interposer hardware" is today.
Probably intentional on apple's part but the old article needs and update
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
219 Comments
Back to Article
meacupla - Tuesday, March 8, 2022 - link
This naming scheme is awful.As if the names weren't already confusing with Pro and Max, now we get Ultra, which is double Max.
So okay, now your Maximum wasn't actually maximum, but half a maximum.
Apple, please stop copying Intel/AMD naming schemes.
trenzterra - Tuesday, March 8, 2022 - link
Perhaps the next in line is the updated Mac Pro with the M1 Max Ultra with perhaps a more performant version with the M1 Pro Max Ultra ;)meacupla - Tuesday, March 8, 2022 - link
I half expect Apple to release the Macbook Studio Pro Max UltraScabies - Wednesday, March 9, 2022 - link
Money's tight, so I'm going to wait for the Pro Max Ultra SEmarkiz - Monday, March 21, 2022 - link
Pro Max Ultra SE Plusbigvlada - Wednesday, March 9, 2022 - link
Well, ATI had Pro Turbo and Ultra Pro, perhaps Apple could use Ultra Pro Turbo. :DAlistair - Tuesday, March 8, 2022 - link
come on, it is super easy, don't be sillyKangal - Wednesday, March 9, 2022 - link
Does this imply that?(iPhone 13 Pro) Max x2 = 1x Ultra (Galaxy S22)
André - Wednesday, March 9, 2022 - link
Yeah, how are we supposed to keep track of M1, M1 Pro, M1 Max and M1 Ultra. So confusing.No no, give me Core i5 124800T, 12400F, 12400, 12490F, 12500T, 12500, 12600T, 12600, 12600KF, 12600K, Core i7 12700T, 12700F, 12700, 12700KF, 12700K and Core i9 12900T, 12900F, 12900, 12900KF and who can forget 12900K? And don't forget the mobile variants Core i5 12450H, 12500H, 12600H, Core i7 12650H, 12700H, 12800H, Core i9 12900H and 12900HK.
That's something we, the consumers, can get behind.
KPOM - Wednesday, March 9, 2022 - link
And some Core i5 chips are faster than some Core i7 chips.melgross - Wednesday, March 9, 2022 - link
Exactly! I don’t know what the beef is.darwinosx - Tuesday, March 15, 2022 - link
There isn't one. Just foolishness.markiz - Monday, March 21, 2022 - link
It's not foolishness, people (still) hold apple to higher standards, because they themselves have set them selves up like that.I mean it does not make too much sense though Like, how ironic is their "think different" campaign now.
varase - Thursday, March 10, 2022 - link
Better yet, a W followed by a random number.CharonPDX - Thursday, March 10, 2022 - link
One difficulty is that they hide different SKUs under the same name. At least with all those ridiculous variations, there *ARE* differences, however minor, between them all.Apple has two M1 chips (7-core GPU and 8-core GPU,) 3 M1 Pro chips (8 or 10-core CPU, 14 or 16-core GPU, with 8+16 being the only disallowed,) 2 M1 Max chips (10-core CPU, 24 or 32-core GPU,) and now two M1 Ultra chips (20-core CPU, 48 or 64-core GPU.)
So nine different CPUs hiding behind only 4 names, with no way to tell the difference other than the longer-than-a-product-number actual description of CPU and GPU cores.
And we don't know about thermal envelope/speed differences between product lines, either. Does the M1 Max in the Mac Studio run faster than an M1 Max in a MacBook Pro? With Intel's product numbers, while they may be obtuse on their face, each one defines a specific spec of cores, speed, extra features, and power envelope.
solipsism - Saturday, March 12, 2022 - link
I can assure you that Apple uses SKUs for all variations of their chips to keep track of their stock. Since they don't sell the chips to others they don't need to offer up confusing SKUs to customers of their machines.Having a designator based on an architecture, scale, with an in-line numerical references to cores, RAM, and storage to be much easier to figure out without having to look up what a long coded SKU means.
Trackster11230 - Wednesday, March 9, 2022 - link
As commented elsewhere, "Ultra" is Latin for "on the far side of, beyond", so it's not really incorrect. But I still agree these naming schemes are silly. Is the Mac Pro going to be the M1 Turbo^2?Kangal - Friday, March 11, 2022 - link
For me, the monikers ascend in this order:Lite > Base > Extra > Plus > Turbo > Pro > Super > Mega > Hyper > Extreme > Ultra > Max
...that makes the most sense/gut-feeling for me.
oryanh - Thursday, March 10, 2022 - link
It's like buying hairspray in the 90's... "Do you want Aqua Net - SuperHold, MegaHold, or UltraHold, or MaxHold"?darwinosx - Tuesday, March 15, 2022 - link
Ridiculously whiny comment and you obviously don't know how AMD names their processors.Iketh - Thursday, March 17, 2022 - link
calling out a whiny comment with a whiny comment, nice!arglborps - Friday, March 25, 2022 - link
But the naming is correct. Max means maximum which is the maximum die size. And Ultra is Latin for "beyond" which means they go beyond maximum die size by having two dies. Linguistically and logically the naming is rock solid.mode_13h - Saturday, March 26, 2022 - link
Interesting. Thanks for sharing!zamroni - Friday, April 22, 2022 - link
So, there will be iphone 14 ultra as well?anonomouse - Tuesday, March 8, 2022 - link
"This means twice as many CPU cores, twice as many GPU cores, twice as many neural engine cores, twice as many LPDDR55 memory channels, and twice as much I/O for peripherals."Really excited to hear about this LPDDR55! Skipping 50 generations is pretty impressive.
trenzterra - Tuesday, March 8, 2022 - link
Did apple just one up the recently announced ucie?meacupla - Tuesday, March 8, 2022 - link
how did you come to that conclusion?michael2k - Wednesday, March 9, 2022 - link
I think the assumption is that the 2.5TB/s Ultrafusion interconnect is higher bandwidth than the UCIe/PCIe5 bandwidth not that I can find information as the chiplet to chiplet bandwidth of UCI; is it PCIe5/CXL that determines the chiplet to chiplet bandwith?They mention 1.3TB/mm, but not how many lanes (it says spare lanes in advanced), but I assume at least 2x, so 16 lanes?
https://www.anandtech.com/show/17288/universal-chi...
trenzterra - Tuesday, March 8, 2022 - link
And anyway, my first thought was... Could this finally end the GPU shortage? Imagine miners snatching up Macs in droves for eth miningIntelUser2000 - Tuesday, March 8, 2022 - link
A search reports the M1 Max getting 10MH/s. It's pretty power efficient using 19 extra watts, but good luck with reaching ROI... ever. Double that you get 20MH/s, which is what you'd get with 6 year old RX 470 at stock settings.caribbeanblue - Monday, March 21, 2022 - link
Keep in mind, afaik the current mining methods on macOS don't really take advantage of Apple's GPU architecture or Metal.ZoZo - Wednesday, March 9, 2022 - link
The GPU shortage is already endingQasar - Thursday, March 10, 2022 - link
oh ? not going by the stock levels of a couple of local stores here, most cards are still out of stock. have havent changed for quite a while now. been keeping an eye on it as i would like to upgrade my GTX 1060 :-)domboy - Thursday, March 10, 2022 - link
That would be hilarious... mining farms made of up these M1 Ultra machines, causing a Mac shortage instead of a GPU shortage.lilkwarrior - Friday, March 11, 2022 - link
Mac Studios are way too expensive for non-creatives to try and use for use cases like mining at scale. It's far more sensible to get a 3090/6900XT until point-of-stake crypto-mining eventuallyIntelUser2000 - Tuesday, March 8, 2022 - link
The performance claims of exceeding RTX 3090 is total nonsense. Maybe in some mobile GFXBench where it's optimized for mobile architectures. Maybe doubling M1 Max can get RTX 3080 Mobile performance.phr3dly - Tuesday, March 8, 2022 - link
Thus far most of Apple's performance claims about its silicon have borne out. Obviously independent testing is important, but IMO they've earned the benefit of the doubt.fishingbait15 - Tuesday, March 8, 2022 - link
That is because Apple has either:A. made general claims without referencing a specific device
B. made specific claims concerning Apple devices running 8th, 9th and 10th gen Intel Core i3, i5, i7 and i9 CPUs.
C. after it was demonstrated that Alder Lake - and some Tiger Lake and AMD - CPUs outperformed the M1, M1 Pro and M1 Max in configurations that cost hundreds less Apple subtly moved the goalposts to "industry leading power per watt" (in large part due to using TSMC's industry leading process that AMD was as much as 2 generations behind and Intel up to 4).
Apple has proven that they can make comparable CPUs to Intel and AMD but not better ones. They have a massive lead with integrated GPUs that Intel is actually starting an entire new research division to close the gap but that is the only area where they have a clear advantage. All the others are areas and metrics chosen by Apple for the purposes of giving them an advantage.
Finally, in addition to being on the latest TSMC process, a huge chunk of the M1's performance is due to replacing RAM with unified memory. Intel and AMD have no interest in copying this directly, but the chiplet idea that Intel has proposed and the rest of the industry (except Apple and Nvidia) has agreed to adopt has the potential to approximate it. It won't integrate the RAM on the SOC proper, but it will connect the RAM to the CPU on a connection that provides performance similar to what Apple is using to connect two M1 chips to make the M1 Ultra.
valuearb - Tuesday, March 8, 2022 - link
Apple has made performance per watt their primary goalposts since the announcement of M1, that was the first and most prominent slide. And I haven't seen Alder Lake outperforming the M1 Max in anything with a battery life over 4 hours.We expect Alder Lake laptops to be cheaper because they'll be made in cheaper build quality laptops without premium design features from vendors stuck with WinPC 2% margins instead of Apples 20% margins. But I haven 't seen any yet.
Calin - Wednesday, March 9, 2022 - link
I haven't checked in a while... however, if you wanted both the performance and the build quality (display quality, metal cases, ...) of the Apple, you usually paid the Apple price (but buying Dells, ...).s.yu - Friday, March 11, 2022 - link
Yeah but almost certainly with touchscreen and 360° flip, and possibly with a digitizer too.Eug - Tuesday, March 8, 2022 - link
I always find it odd when people argue "Apple is only beating AMD/Intel because they have implemented this technology and have that other advantage".Alistair - Tuesday, March 8, 2022 - link
Alder Lake wasn't out when the M1 Max released, yeeshOreoCookie - Tuesday, March 8, 2022 - link
With regards to your comments:A. Actually, if you look at the footnotes of https://www.apple.com/mac-studio/">Apple's Mac Studio page they give the specs of the systems they benchmarked the M1 Max and M1 Ultra against. Ditto for the slides they used in the keynotes, the system specs are given on the bottom right. One of the systems they benchmark it against runs a Core i9-12900K, i. e. the fastest consumer-level Alderlake processor.
B. And they do so for Alderlake, i. e. 12th gen Core processors, too.
C. Performance-per-watt and its effect, battery life, were selling points from the very beginning. When the M1 Max was benchmarked, AT remarked that only a high-end desktop part with a much, *much* higher power consumption could beat it. Of course, time moves on and processors with much higher TDP have beaten M1-class processors in various benchmarks.
To claim Apple is “conveniently” switching to performance-per-watt as their yard stick is inaccurate. In fact, it was the reason they switched to Intel over 15 years ago. Intel knows it is seriously behind: in a leaked company slide deck, Intel employees were claiming that they intend to beat the M1's efficiency in 2024.
Lastly, you write that Apple's advantage is due to replacing RAM with unified memory. That's where some of its advantage comes from, yes. So what? This seems like a very good idea to me. Previously, the big downside of an integrated GPU was that memory bandwidth was scarce: CPUs used to have a much lower memory bandwidth to begin width, and the GPU gobbins had to share the meager bandwidth with the CPU. Hence, the association of integrated graphics = slow graphics. Once you lift the bandwidth bottleneck and allow for high-enough clockspeeds, you see that there is no performance compromise, but the opposite rather.
mdriftmeyer - Wednesday, March 9, 2022 - link
The claims are erroneous. They cite no specific tests on those early claims. Just Up to 3.4 times CPU performance. Of what? Well, we won't tell you.It's the same crap Intel pulls.
OreoCookie - Wednesday, March 9, 2022 - link
Yes and no. Apple has a history of relatively accurate performance claims. They are typically in line with a broader mix of benchmarks and have been verified here on Anandtech. In recent years that hasn’t always been the case with Intel. For this reason, I’d take that as a ballpark figure and am looking forward to the in-depth review here. :-)AceMcLoud - Wednesday, March 9, 2022 - link
"Apple subtly moved the goalposts to "industry leading power per watt" "That is a lie, it's one of the main points they have been highlighting from day 1.
Blastdoor - Wednesday, March 9, 2022 - link
And by ‘Day1’ you could easily mean day 1 of the transition from PPC to intel. It was IBM’s inability to fab a G5 that could go in a PowerBook that pushed apple to intel in the first place.The M (and A) chips reflect apple priorities going back 20 years. In addition to performance/watt being a long-standing priority, specialized SIMD units (remember altivec?) and using the GPU performance for non-gaming purposes has also been a long time priority.
Basically apple had made the chip they always wanted and that intel could not make for them.
ingwe - Wednesday, March 9, 2022 - link
Yeah I think this is absolutely the take. Apple has made a huge part of its laptop segment since moving to intel about superior battery life. Only tangentially related, but I feel like they haven't pursued that on iPhone to nearly the same extent.Woochifer - Thursday, March 10, 2022 - link
Yes and no on the iPhone. They have pursued power efficiency on their mobile devices, but more as a tradeoff for making those devices thinner. Generally, the battery capacity on iPhones is smaller than equivalent Android models, so the efficiencies with the SoC and OS optimizations have been deployed more in support of that tradeoff.Prebengh - Wednesday, March 9, 2022 - link
If you care to look at the Apple product page for the Mac Studio you can see how Apple carried out the tests for the comparisons.melgross - Wednesday, March 9, 2022 - link
They actually have been comparing them to specific devices. The devices are listed at the bottom.whatthe123 - Tuesday, March 8, 2022 - link
what are you talking about? they claim nearly 2x the performance of competitors even with competitors pumping up power draw, and better GPU performance than the 3080 laptop edition, yet they've never hit those numbers.Alistair - Tuesday, March 8, 2022 - link
Shadow of the Tomb Raider is already faster on M1 Max than on RTX 3080 mobile... "never hit those numbers" is what you are saying? Huh?whatthe123 - Wednesday, March 9, 2022 - link
what? https://www.anandtech.com/show/17024/apple-m1-max-...Alistair - Wednesday, March 9, 2022 - link
Depends on the TDP of the 3080 laptop you are comparing. See this:https://www.youtube.com/watch?v=WyB3G_q33N4
mattbe - Wednesday, March 9, 2022 - link
This Youtube video didn’t even run a proper benchmark. If you are going to make bold statements like it’s faster than the 3080 mobile in Tomb Raider, please actually back it up with a proper source where some semblance of a benchmark is conducted.In the mean time, I would much rather rely on Anandtech’s results and Hardware Unboxed.
https://youtu.be/0sWIrp1XOKM
HBU tested the M1 Pro, and you can just double that performance to see that it still isn’t remotely close to the 3080.
techconc - Wednesday, March 9, 2022 - link
Tomb Raider isn’t even a native app on Apple Silicon. Try comparing native apps. Also, GFXBench validates Apple’s claims.bifford - Monday, March 14, 2022 - link
The Anandtech article you quote concludes that the M1 "chips here aren’t only able to outclass any competitor laptop design, but also competes against the best desktop systems out there"whatthe123 - Wednesday, March 9, 2022 - link
also strangely you are the only person on the internet claiming shadow of the tomb raider runs faster on the max. even apple hasn't made that claim.Alistair - Wednesday, March 9, 2022 - link
no, i just watched a video of a 3080 mobile laptop get beaten by the M1 Max, it depends on the TDP of the GPU in that laptop of course, check out Tally Ho, the video link is aboveBillBear - Wednesday, March 9, 2022 - link
Shadow of the Tomb Raider makes a poor comparison point, since it's running x86 code through Rosetta and using OpenGL instead of Metal.World of Warcraft would be an example of a game that is both ARM native and targeted to Metal so at least you are seeing the full performance of the hardware.
Sort of like an how an x86 based XBox could run some of the older games written for PowerPC based XBox consoles, but emulation always slows things down.
shikibyakko - Thursday, March 10, 2022 - link
I mean, most independent testers were unable to really confirm much of the claims of the M1 Pro vs M1 Max, specially because there is really not that many applications that make use of things like Metal 2 to really compare those claims.The performance of the M1 Pro and M1 Max, is impressive in many factors, specially for battery life, but outside of that, they are still waaaaaaay too far away from high end desktop grade performance, not to mention that they are not even close to workstation/server grade CPUs.
Maybe the Ultra will be able to finally close that gap with desktop grade performance, but, as nvidia well knows at least for GPUs, adding double the GPUs dies does increase performance by double, but also introduces double the problems to the point that Nvidia has mostly given up on that concept for a while now.
I think the fact that they have to pull out this trick of making 2 chips in order to finally do a desktop shows one of the main drawbacks of using ARM64 as an x86-64 replacement in this space.
Lets just see how it goes, but I'm very skeptical of their wild claims.
fishingbait15 - Tuesday, March 8, 2022 - link
Exceeding RTX 3090? Sure, total nonsense. RIVALING RTX 3090 ON AN INTEGRATED GPU? The opposite of total nonsense considering that Intel's competitor - the Iris Xe - doesn't even surpass an MX350, let alone an RTX 3090.meacupla - Tuesday, March 8, 2022 - link
While I think RTX 3090 is the faster silicon, I wouldn't knock integrated GPU performance.A properly designed integrated GPU would feature unified memory, which allows for zero-copy, and removes a lot of bandwidth bottlenecks inherent with Ram stick based CPUs joined with PCIe based GPUs.
PS5 is a good example of fast integrated GPU with unified memory. It's just that PS5's GPU is bound by cost constraints.
Alistair - Tuesday, March 8, 2022 - link
the 3090 is not faster, why would you believe that? you have no basis for thatRezurecta - Tuesday, March 8, 2022 - link
It doesn't matter if it's an Integrated GPU or not.The amount of transistors they are dedicating to the GPU is more than what the 3090 uses so it makes sense.
The other difference, as Ryan states, is that since Apple goes extremely wide, they can jam all these transistors together and run them at low speeds so they use much less power! One of the big reasons that AMD can't put a 6900XT on a Ryzen 9 die is because they run high clock speeds on both and the heat would be unmanageable.
Alistair - Tuesday, March 8, 2022 - link
It's one hundred percent accurate, why can't you believe it. The computer costs $5000 for god's sake. The RTX 3090 is beaten by the $1000 6900 XT in comparison.blppt - Wednesday, March 9, 2022 - link
"The RTX 3090 is beaten by the $1000 6900 XT in comparison."Definitely not when raytracing enters the equation.
Alistair - Wednesday, March 9, 2022 - link
and nobody should use ray tracing, ray tracing is total crap, I play at 1440p and 120fps, and you can't do that with any modern game with ray tracing onOxford Guy - Wednesday, March 9, 2022 - link
120 FPS is irrelevant for some genres of games that benefit from rich graphics, such as turn-based RPG. Those require few frames but benefit from graphical richness. Various simulation games do not require a high frame rate.I find it humorous and unfortunate that the mentality of PC gamers continues to be locked into FPS-style games. It’s an entire genre of art; there is a lot more possible.
blppt - Wednesday, March 9, 2022 - link
While I agree that so far the overall RT performance has not been up to snuff even with this generation of cards (particularly the 6900XT), there are always a portion of gamers who prefer eye candy to throwing 200fps at the screen. If there weren't, Nvidia and AMD wouldn't have bothered supporting that feature.mattbe - Wednesday, March 9, 2022 - link
Maybe in certain games, but not in any meta analysis.I am convinced that you are a troll at this point. This is pretty sad.
Alistair - Tuesday, March 8, 2022 - link
also don't forget the 3090 is almost 2 years old now, Apple is ahead now, wait for the 4090 to beat Apple, that is the cycle, don't claim ignoranceOreoCookie - Tuesday, March 8, 2022 - link
You seem stuck on the association integrated graphics = slow graphics. Current-gen consoles already proved this wrong, and finally with the M1 Pro and M1 Max we know that this also works for other platforms. Since Anandtech has already benchmarked the M1 Max, we know what to expect for the Ultra. In gaming benchmarks Apple's chips aren't doing so great, but that is down to drivers. In productivity applications the M1 Max (not the Ultra) already got into the same spheres as fast discrete desktop graphics. Plus, the RTX3090 is getting old and is built on an older process, so Apple's claims aren't surprising.robotManThingy - Wednesday, March 9, 2022 - link
Being "integrated" is now an advantage. I fail to see how GPUs separated from the CPUs and memory are somehow going to be faster than when they are all bundled together using a common memory space.name99 - Wednesday, March 9, 2022 - link
Spoken like a true gamer.You do realize that the world is larger than games, right?
A HUGE aspect of these chips is that they provide a much larger memory capacity for the GPU than either the 3090 (24G) or even a maxed out 48GB Quadro. Which is of immense importance to, for example, people playing with large neural nets...
Yes yes, we get it, Apple is not the preferred platform for playing games. This is not news, is not interesting, and will not change soon. What matters is that it's the preferred for people engaged in tasks other than playing games...
Alistair - Tuesday, March 8, 2022 - link
no, their performance claims have been spot on so far, just remember most games are not programmed well for Macthe M1 Max already beat the RTX 3080 in Shadow of the Tomb Raider, so this new chip will beat the 3090 by a large margin in some games, and be much less in others
Alistair - Wednesday, March 9, 2022 - link
(just to be clear I meant it beats the 3080 mobile at a limited TDP like 100W or less, the 3090 is a lot faster, so the M1 Ultra being twice as fast as the M1 Max is necessary to equal the 3090)halo37253 - Wednesday, March 9, 2022 - link
I'm sorry but the M1 Max Struggles to even compete with a RTX 3060 mobile in games.There has been plenty of tests done on this. The M1 Max GPU is only impressive in synthetic benchmarks. Doubling the cores is not going to help, it will still be slower than a 3080 mobile in actual workloads.
People are quick to try and point at drivers as the reason for low gaming performance is drivers or API calls. But this is just BS, these are not gaming focused GPUs. Back in the day AMD's GCN arch was capable of a good deal higher synthetic benchmarks than nvidia chips while performing around the same game wise.
The M1 is not impressive from a GPU standpoint, and couldn't even compete with something like a steamdeck on performance per watt even given the massive node advantage. The only area M1 is impressive is CPU and Hardware offloading. With the Level of integration, nearly everything being on the package allows for some pretty awesome performance per watt. Add on the massive Node Advantage, AMD nor Intel has any plans to go down that path.
Currently the money is on the Data Center and having access to many PCI-E Lanes for fast SSD Storage. Who knows when we'll get the Mac Pro M1 refresh, as the PCI-E Lanes just isn't there. Maybe Apple will address with with Future M2 Chips, but who knows. Seems like Mac Pro may again go away in favor of Mac Studio. And again back to the Thuderbolt addons.
techconc - Wednesday, March 9, 2022 - link
Are you comparing native games or games running under Intel emulation? Benchmarks have been updated to run native. Everyone likes to compare Tomb Raider for some reason and that game is NOT native for Apple Silicon.OreoCookie - Tuesday, March 8, 2022 - link
Ryan is literally one of *the* experts in the world on GPU benchmarking, and if he says that the performance claims are reasonable (assuming everything works as advertised), then I'd take his word for it.Looking at benchmarks for the M1 Max, we already have a good idea how fast Apple's Mac GPUs are and since graphics is an embarrassingly parallel problem, we know it scales well as you increase the number of cores. Furthermore, Apple has a good track record with their performance claims.
Jumangi - Wednesday, March 9, 2022 - link
I don't know why Apple does these comparisons, nor fanboys either. These Mac are workstations not gaming machines. These attempts at comparisons to something like a RTX GPU are pointless.name99 - Wednesday, March 9, 2022 - link
The comparison is to give people something to latch onto, to compare against something that (some segment of the population) are familiar with.That's all.
The point is not to give detailed comparison specs because, like you say, the actual target buyers do not care about games (or for that matter Cinebench); hence the vague graphs -- all that matters is to let potentials buyers know that "this is the approximate performance level; if that's what you're buying today, take a detailed look at what we have before writing that check".
OreoCookie - Wednesday, March 9, 2022 - link
There are applications that rely on GPU compute, and these benchmarks do give an indication of how performance is scaling. “Consumer-level” GPUs are frequently used for GPU compute tasks, too, so I think the comparison is relevant and fair.And I completely agree with you that gaming is not at the center, Apple’s drivers are not optimized for gaming, so in many cases framerates are below what you’d expect from the hardware in many cases.
Samus - Wednesday, March 9, 2022 - link
Why would it be optimized for a mobile GFXBench or "mobile architectures?" This is a workstation chip, and as such the GPU will be optimized for OpenGL libraries much like AMD FireGL and nVidia Quadro GPU's are.People need to stop associating ARM and RISC architectures in general with mobile devices. Its important to understand a mobile GPU doesn't do anything differently than a traditional nVidia or AMD GPU, but everyone has a vastly different approach to architecture - mostly because of individual IP. Apple took care of that buying Imagination to protect themselves with a patent portfolio they can build on.
mattbe - Wednesday, March 9, 2022 - link
Because the SOC is based on Tile-Based deferred rendering.. it’s more efficient in some ways but worse in others. It’s also more power efficient, which is why it’s used in mobile hardware.The M1 and it’s variants, as with other Apple SOCs, use TBDR.
gfxbench is optimized for TBDR. A TBDR based GPU would suffer a significant performance hit when you have post processing effects compared to GPUs that don’t use it.
techconc - Wednesday, March 9, 2022 - link
Explain how TBDR is worse in ANY way. Explain why you think GFXBench is optimized for TBDR.omi-kun - Thursday, March 10, 2022 - link
Actually, TBDR based GPU would have an advantage in post processing because of their use of tile buffers to cache textures and color buffers for only the current tile, as opposed to immediate based GPUs that can't guarantee all the current data will remain resident on chip due to cache backing and cache contention as a result of multiple SMs competing for limited cache. RDNA2 mitigates this with a much larger cache. But a tile buffer is still more efficient energy wise - smaller buffer closer to the cores, no need to worry about cache conflict.mode_13h - Sunday, March 20, 2022 - link
> Tile-Based deferred renderingI think the "R" is for rasterization.
Nvidia has been doing this since Maxwell. That's how they managed to stay competitive with Fiji (i.e. "Fury"), which significantly out-classed it on raw specs.
AMD has been playing catch-up, here. They added a hardware feature called DSBR (Draw Stream Binning Rasterizer) to facilitate TBDR in Vega, although it remained unused until late in Vega's product cycle. When finally enabled, I think I read it was good for maybe a 10% performance boost.
photovirus - Monday, March 21, 2022 - link
You're correct that tile-based rasterization is used by many vendors. However, it's only part of rendering pipeline (e. g. https://www.techspot.com/article/1888-how-to-3d-re... ).With Apple, R stands for rendering. Their whole pipeline is tile-based. You can watch WWDC videos on how their GPUs work, they've made several in 2020 and 2021.
mode_13h - Tuesday, March 22, 2022 - link
> With Apple, R stands for rendering.Confirmed.
https://developer.apple.com/documentation/metal/re...
Without digging into that, it seems a bit silly to say "rendering" is deferred, because the entire process is collectively "rendering". Perhaps they simply wanted to emphasize that more than rasterization was deferred.
> Their whole pipeline is tile-based.
It's basically what Imagination Tech used to do. They were tile-based since the early days.
caribbeanblue - Monday, March 21, 2022 - link
"It's optimized for mobile architectures" Proof? Source? Just because it runs on mobile devices doesn't mean it's a "mobile benchmark".Rezurecta - Tuesday, March 8, 2022 - link
Definitely super impressive by Apple. I'll wait for benchmarks but what Apple is doing will hopefully push the rest of the industry to have better laptops and power efficient designs!Apple likes to fudge their numbers and make it seem like magic, but its hard work by engineers and process advantages that Apple holds.
Dolda2000 - Tuesday, March 8, 2022 - link
Welp, I hadn't guessed we'd see the first MCM GPU from Apple, but apparently here we are. It will be interesting to see how it actually performs in practice, but it seems that most review houses have been having trouble finding like-for-like GPU workloads to compare between the Apple ecosystem and anything else.Alistair - Wednesday, March 9, 2022 - link
yeap Apple has a 6 month lead here, but they'll fall way behind by this fall and then not update the studio for 2 years... same old same oldrcfa - Wednesday, March 30, 2022 - link
You assume that Apple sticks to the same product cycles as when they were bogged down by the lackluster intel performance increases.That Apple can keep up a steady 12-18 month product cycle where it matters and they have full control should be obvious from all their i(Pad)OS based devices.
So, I’d bet, your assumptions are not only wron, but also a psychological defense mechanism, because the prospect of Apple dominating desktop performance is a concept you haven’t made peace with, yet. You will, in time, have little choice…
mode_13h - Thursday, March 31, 2022 - link
> psychological defense mechanismFor an otherwise reasonable post, this is an unnecessary hit below the belt. Let's try and stick to facts and logic, while accepting that reasonable people can disagree and each is entitled to their own opinion.
Eug - Tuesday, March 8, 2022 - link
Whether or not it truly beats RTX 3090 is kinda besides the point IMO. That is probably not as important to Apple as having ALL of the M1 Ultra machines with the exact same very high GPU performance, at a much lower power budget. And don't forget that this is effectively a small form factor design. The top end Mac Pro hasn't even been announced yet. I am guessing that will be announced in June at WWDC.valuearb - Tuesday, March 8, 2022 - link
The Mac Studio is a iittle pocket monster that I'm sure will be super popular with creative professionals. The $4k version has both a more powerful CPU and GPU than a $15,000+ Xeon Mac Pro.But it doesn't have any expandability outside all those (admittedly super fast) thunderbolt ports and doesn't have user upgradable RAM or ECC RAM. What is even more interesting will be the next Mac Pro, which is rumored to have dual sockets for two M1 Ultras, along with expansion slots and a big chassis. Will it support expandable ECC RAM? It's possible they might provide expansion through CPU swaps. Start with a "low end" single 64 Gb Ultra, later pair it with a 128Gb high end Ultra when you need more power, and eventually when it's cheap in a few years swap out the low end Ultra for another high end one.
Silver5urfer - Tuesday, March 8, 2022 - link
Never going to happen. Apple will not engineer an LGA socket for their uber niche market, this thing won't even be bought by masses. Look at the price tag, $1500 base for pathetic M1 Max. And 4K for the real upgrade and to beat a 3090 you will need to add 2K more to make it 6K and then even more for more storage and memory.Mac Pro will not come now, I think it might be axed or maybe Apple just for bragging rights will do one for M2 series. As you mentioned this cannot be a workstation class machine. Its too niche and too expensive plus everything is soldered.
andynormancx - Wednesday, March 9, 2022 - link
Apple literally announced at the end of the presentation that the Apple Silicon based Mac Pro is still coming.Lavkesh - Wednesday, March 9, 2022 - link
Pathetic M1 Max? You are literally butt hurt fanboy of somethingt.s - Tuesday, March 15, 2022 - link
Excuse me. He or you the fanboy here?Luis Alejandro Masanti - Tuesday, March 8, 2022 - link
It seems to me that they need ‘two different chips’: the original M1 Max and a ‘mirror copy’… so the can connect the side wires in the correct order.If we have one chip with —say— the connector over the right side and the first wire is at the top… when we rotate the other chip so the connector is on the left… the first wire would be in the bottom…
…except that the ‘bridge’ inverts the 10.000 wires.
Which solution would be easier to do?
A5 - Wednesday, March 9, 2022 - link
From their die shots it is just rotated 180*.The Von Matrices - Thursday, March 10, 2022 - link
About half of the pads are for transmitting data and the other half are for receiving data. You don't want to connect the same pads together otherwise you would be connecting transmit pads to other transmit pads and receive pads to other receive pads.Usually, the chips alternate transmit/receive pads (e.g. pin 1 transmit pad 2 receive.... pin 1023 transmit, pin 1024 receive) and then when the chip is flipped around, one chip's transmit pads line up with the other chip's receive pads (e.g. pin 1 transmit -> pin 1024 receive, pin 1023 transmit -> pin 2 receive.)
Silver5urfer - Tuesday, March 8, 2022 - link
Everything falls apart for singing praises for this garbage design which Apple excels at best - BGA solder.This processor is having $1500 price tag for a basic M1 Max SOC and with $4000 cash you get the Ultra upgrade, but the DRAM is unified so the max memory for 4K is 64GB, and that will eat into the GPU memory, okay it's huge and very fast. Still it's nothing ground shattering since how Apple's M1 Max and Pro got slaughtered in dGPU tests last time at Anandtech, I remember how they lose to a crappy mobile GPU which is 3080L / 2080 Desktop class on Aztec Ruins and in SOTTR it crawls, at 3060L lol. Awful garbage.
Now we have this omega marketed halo product which commands such a high price tag with similar expectations and for the dGPU performance it will again be Apple OS based vs the GPU rendering power, and since ARM processors use specialized blocks and this one has more blocks for decode and encode it will be faster as Apple claims however the advantage is very thinly spread due to it's niche nature.
Now since it's already mentioned how Apple SOC is massive in area and the Transistor count, and the most funny part is look at the CPU and GPU performance 12900K and 3090, now they say it's almost at that level for the most part and GPU is where they beat and ofc it's all similar performance only advantage where they beat is Power consumption. So ultimately pay $6000 USD to have a BGA soldered fancy looking box which will have exactly similar performance to a latest PC LMAO. And ? We are supposed to take this rubbish seriously ? Yeah I guess.
Meanwhile Z690 or any Mainstream mobo has PCIE expansion slots for vitually anything, and a lot of I/O instead of the BS Thunderbolt ports which again needs adapters to have multiple ports extra money needed. And the worst part is even the Storage is soldered. Like pay 4K + extra 1-2K for 4TB, for that amount I could buy Intel Optane P5800X which will shred this pile of junk to ash and outlive this silicon, OR I could simply buy FireCuda NVMe SSD which is made of Micron's latest flash 196L and gets me insane TBW endurance and performance, ofc I can keep it cool because I use a Z690 motherboard.
As for Workstation class, this product is a giant failure. You cannot run any sort of VMWare Hypervisor home server class software OS, there's simply no way this junk will support that, on top you have to spend a ton of cash. Why not simply get an office decommissioned XEON processor and a good PowerEdge / ProLiant HPE chassis and add HBA SAS expander and blow the roof out of the RAID ? Or say for eg buy the new upcoming Threadripper series which is a HEDT machine and will churn everything you throw at it and will live super long long time and on top you can get a TON of DRAM memory plus the massive I/O bandwidth ?
Yep, I will forget everything and simply simp for this junk company and their bragging of TSMC first hand product access. Wake us up when AMD and Intel and Nvidia get TSMC 5N lithography access and let's see the M1 Ultra get slaughtered.
Hmmmokay - Wednesday, March 9, 2022 - link
That’s very aggressive, for no reason. People are excited by new technology. If you aren’t excited, that’s fine.Dolda2000 - Wednesday, March 9, 2022 - link
I don't think it's entirely uncalled for. While Apple has some technology that I'd normally be quite excited about, the level of excitement is quickly checked by it being... Apple. A $4,000 system that is completely nonupgradable, nonmodular and unrepairable, with even the SSD being soldered? The only way you can get behind that in good conscience is it were at least twice as fast as the next best competitor, but it really is not. And with all their grandstanding at every turn about their commitment to "the environment" while still maintaining absolute resistance to any form of independent repair, I'm not sure calling them a "junk company" is at all unfair.It's a great big shame, really. There's a lot of things in the M1 series that I'd like for myself, but it only being available in Apple products just kills it for me.
andrewaggb - Wednesday, March 9, 2022 - link
I like where a lot of the M1 is going. I'm curious how much is better design, how much is just having an enormous transistor budget, how much is better manufacturing, and how much is false marketing.For example, comparisons to nvidia products made on samsung's process aren't fair from a power consumption point of view at all. The samsung process is way worse and the 3090 wasn't tuned for power consumption, it's a massive card with a huge heatsink and cooler.
The massive memory bandwidth and faster integrated graphics is something I've often wondered why AMD doesn't pursue in a laptop/desktop form (like a ps4/ps5) as clearly they know how. It's not clear to me why Apple is doing it first but good for them. Personally I would probably be willing to trade off upgradable ram for better performance if they don't price gauge you for it. For SSD's I can't see any valid reason to make them non-upgradeable as PCIE4/5 isn't holding them way back. To date I've never seen a vendor that doesn't price gauge on RAM and SSD upgrades so I'm not hopeful in this regard.
I also see some misreporting, eg there's an article on Toms claiming a 3990x threadripper is the same on geekbench but if you look at the linux (not windows) results for the 3990x it's 40% faster. There are also claims of 3090ish performance but then independent testing shows way lower results most of the time. And something like ethereum mining is dreadfully slow on the m1 variants by all accounts, etc.
At the end of the day I still think it's probably the best laptop chip you can get but I hope we see some more competition.
mode_13h - Sunday, March 20, 2022 - link
> massive memory bandwidth and faster integrated graphics is something I've> often wondered why AMD doesn't pursue in a laptop/desktop form
$$$
The consoles and this Apple Studio Pro are vertically-integrated, which means they're not stacking margins on top of margins. Furthermore, MS and Sony lose money on the console hardware and make it up in software royalties and subscription services.
If AMD makes a monster APU, it's going to be expensive to the point where it'd be a niche product and most in that price range would prefer to get just a CPU and dGPU.
Worse, it would limit you to non-upgradable memory, since you can't get the needed bandwidth with DIMMs.
> clearly they know how.
Right. They *do* know how, and haven't seen fit to do it for the past 10 years since they started making console APUs. That's your clearest sign that it hasn't been viable for the mainstream market thus far, because they've been keen to play up their GPU advantage over that entire time.
valuearb - Sunday, March 20, 2022 - link
SSD is replaceable. Macs last forever. It's got a ton of super high speed thunderbolt ports for all the external storage and devices you want. And its hands down the fastest video and graphics production system you can get under $5k.mode_13h - Monday, March 21, 2022 - link
It'll be interesting to see how long that LPDDR5 lasts. DRAM does wear out with use and all the heat it's going to be feeling from the CPU won't help!Dolda2000 - Saturday, April 2, 2022 - link
>SSD is replaceableOnly with original Apple parts (since it's not a standard NVMe drive), which not only means that you cannot upgrade it with third-party options, but also that you can't take the SSD out of an otherwise potentially faulty system and get the contents out of it in another system.
>It's got a ton of super high speed thunderbolt ports for all the external storage and devices you want.
What if I don't want external storage, but just larger internal storage? Or RAIDed internal storage? Or more memory? Or a 32-core GPU with only a 10-core CPU? Or to upgrade my CPU (or the WiFi adapter) three years down the line? And so on and so forth.
>And its hands down the fastest video and graphics production system you can get under $5k.
Only because of the media engines, and there's absolutely no reason they couldn't be on something like a PCIe card.
>Macs last forever
At this point I hope you're trolling.
rcfa - Wednesday, March 30, 2022 - link
“Repairability” is the biggest cause for necessitating repairs.It is much better for the environment to toss one and recycle one in e.g. every 10’000 units than to toss/recycle zero units, but have to repair one in 1’000.
Apple has the volume to make these statistics, and they optimize globally not based on the emotions of a single affected consumer.
Reliably of glued-down, soldered-in-place parts is much higher than that of removable, socketed parts.
While the latter can be repaired, stock kereping, packaging, shipping of repair parts has a significant environmental footprint, too; and multiplied by the higher failure rate, the environment takes a beating compared to the non-repairable approach.
mode_13h - Thursday, March 31, 2022 - link
There are valid points on either side. Neither approach is going to be right for everyone, and some people just value the expandability of a machine that's designed to be upgradable.I have a workstation with a NVMe SSD in a PCIe slot. It has no M.2 slots, but I have enough PCIe lanes that I could add a PCIe card when they came along. I like being able to do that sort of thing. I also upgraded the GPU in this machine, which was a huge benefit, though I also had to swap around PSUs when I did. I was also glad to have that option. Another mid-life upgrade I typically do is to double a machine's RAM. Sometimes, I've even done 2 RAM upgrades, doubling capacity each time.
At work, I've upgraded machines' CPUs, RAM, and GPUs to fulfill project needs.
So, as much as you might favor a certain approach, you really can't say it's right for everyone and every purpose.
Dolda2000 - Saturday, April 2, 2022 - link
>Reliably of glued-down, soldered-in-place parts is much higher than that of removable, socketed parts.Do you have a citation for that claim? It seems immediately false to my experience. Extremely rarely have I had such components as socketed CPUs, GPUs, RAM or SSDs fail, and even less so in a way that is traceable to them being socketed instead of soldered. Intrinsic failures appear to be much more common with such things as capacitors or other simple components that are merely part of something like a motherboard or PSU, and in reality failures are much more commonly caused by some sort of user error or other external factor (liquid damage, dropping things, power surges, and so on), which would affect a monolithic system just as much as a modular system.
>It is much better for the environment to toss one and recycle one in e.g. every 10’000 units than to toss/recycle zero units, but have to repair one in 1’000.
To begin with, I don't believe that's true, unless you have some citation for that as well. It seems obvious to the naked eye that a proper and limited repair has a significantly more limited impact. Changing a bad cap or cleaning up liquid damage is literally nothing compared to recycling and assembling a whole system. I also don't believe that modular systems fail more than an order of magnitude more often than monolithic systems. Nothing in my experience tells me that would be the case. See above, I guess.
>Apple has the volume to make these statistics, and they optimize globally not based on the emotions of a single affected consumer.
Is your argument actually just "trust Apple, they're just such good guys, they would never do anything to screw over the customers in order to improve their bottom line, it just looks that way to you because they are optimizing for global variables"?
mdriftmeyer - Wednesday, March 9, 2022 - link
Not that I'm raving about the Studio but you do realize with a small form factor NUC design you would slave a Thunderbolt Hub off the Studio, then slave off a NAS and be done with it.I agree the Zen 4 will turn heads on the 5N process this Fall.
kwohlt - Monday, March 14, 2022 - link
Those are certainly all words." Like pay 4K + extra 1-2K for 4TB" Doesn't seem entirely unreasonable. Go to Lenovo's website and configure a P620 workstation and see how you can pay $1K for 2TB of NVME. $1K per TB is how much I expect to pay for enterprise storage solutions anyways.
"you cannot run any sort of VMWare Hypervisor home server class software OS". Lol who cares? We run our hypervisors on our datacenters on proper servers, not workstations that live on employee's desks.
"Why not simply get an office decommissioned XEON processor and a good PowerEdge / ProLiant HPE chassis and add HBA SAS expander and blow the roof out of the RAID"
Buy used, slower hardware instead because it's cheaper?
"Or say for eg buy the new upcoming Threadripper series which is a HEDT machine and will churn everything you throw at it and will live super long long time and on top you can get a TON of DRAM memory plus the massive I/O bandwidth ?"
I guess that's always an option, but expect to pay a lot. Lenovo still charges $4500 to upgrade to a W3975WX, and you're still left with significantly less VRAM no matter what discreet card you go with, and consuming multiple times more power, from a much larger device.
Not sure why you're hating so much - M1 Ultra and Mac Studio are very price competitive and offer incredible perf/watt, and it's not even the highest tier M1 series SoC that's going to be released.
t.s - Tuesday, March 15, 2022 - link
I think everyone sane. And their mother, grandma, plus their dog, want something repairable and upgradeable. Not some premium junk-that-when-break-go-straight-to-bin like these M1 what-what-ever. Soldered RAM still ok (although you must pay some hundreds bucks for 2xRAM size, lol). But soldered SSD?! Damn Apple and their fanboy that said it's OK configuring something like that.Same with x86 soldered and unrepairable junk. Like most of x86 chromebook and MBA-wannable-ultrabook.
valuearb - Sunday, March 20, 2022 - link
Macs last forever and depreciate slowly. When I want more performance than my M1 air can offer, I'll sell it (for still a very good price) on Ebay and buy new.valuearb - Sunday, March 20, 2022 - link
Your butt pain is very tangible.mode_13h - Monday, March 21, 2022 - link
Oh, he's like that.TechDoc - Wednesday, March 9, 2022 - link
This is a huge step forward for chiplet technology. Another step forward would be edge-to-edge interconnects on top of the 2.5D interconnects. Those interconnects are even faster, and can support multiple signal types, such as analog/digital/optical/power all on the same edge. An Indiana company is developing that: www.indianaic.com.usiname - Wednesday, March 9, 2022 - link
With 114B transistors you can build 6900XT+RTX3090+3*5950xAlso clocked low and using 5nm TSMC. it would be ridiculous not to be more effective, but that ladies and gentlemen is how to laste transistors on closed system in time when transistors are not enough for everyone
usiname - Wednesday, March 9, 2022 - link
Ohhh also... 1000kWh less per year =>100-200$ less per year. Good luck saving money with this useless product.techconc - Wednesday, March 9, 2022 - link
You’re comparing transistor count of a gpu to an entire SoC. Likewise, it’s clear that don’t even understand what you’re complaining about.mannyvel - Wednesday, March 9, 2022 - link
Funny how many people are saying "unbelievable." That's what people said last year, when a juiced up mobile chip kicked the behinds of every desktop chip except for the highest end ones. When it comes to reality, your belief is not required.bottlething - Wednesday, March 9, 2022 - link
Curious about this unimaginable irrelevant tidbit:Weight (M1 Max): 5.9 pounds (2.7 kg)2
Weight (M1 Ultra): 7.9 pounds (3.6 kg)2
Glaurung - Wednesday, March 9, 2022 - link
Bigger heatsink on the Ultra, I assume.Luis Alejandro Masanti - Wednesday, March 9, 2022 - link
Quick answer: Each M1 Max chip weights 2 pounds!Not so quick answer: The fan and power supply can be bigger and heavier too.
BillBear - Thursday, March 10, 2022 - link
The heatsink on the Max is aluminum.The heatsink on the Ultra is copper.
Copper weighs more than aluminum, but is better at moving heat, which is why coolers often have copper heat pipes even when they have aluminum radiators.
SydneyBlue120d - Wednesday, March 9, 2022 - link
And still no Autocad for M1...Alistair - Wednesday, March 9, 2022 - link
they are sure taking their time...MC_Kubbe - Wednesday, March 9, 2022 - link
Actually Apple is informing us the GPU is 21 TeraFLOPS, so yeah - it's not more powerful than RTX 3090.Butterfish - Wednesday, March 9, 2022 - link
Teraflops numbers can’t be compared across different architectures. It is a simple figure that are calculated by timing frequency and the numbers of ALUs together. How will are these ALUs been used to be turn into real world performance was not reflected by these figures.Before the introduction of NVIDIA’s Ampere architecture. NVIDIA used to have a slight lead with relative performance per TFLOPs. Their 1000s series GPUs can outperform AMD competitor at the time despite having lower TFLOPs. However, with introduction of Ampere comes big architecture changes. NVIDIA decided to double the amounts of floating point ALUs by having the newly added ALUs share the pipeline with integer ALUs. This mean half of floating point ALUs on Ampere GPU have to choose to compute either floating point OR integer calculation, but not both simultaneously. This mean not all floating point ALUs can be utilized at any given time if integer operation was required. Also other bottleneck might exist in the architecture to properly feed twice the amount of floating points units. Regardless, NVIDIA didn’t achieved doubling the performance of traditional raster or shader performance with this modification, this mean now NVIDIA’s TFLOPs number are highly inflated compared to relative performance.
AMD’s 6900 XT has 23.04 TFLOPs, but it can trade blow on raster performance with RTX 3090 despite the later having 35.58 TFLOPs.
techconc - Wednesday, March 9, 2022 - link
It looks like someone doesn’t understand the difference between theoretical and actual performance. Integrated systems come much closer to theoretical performance than discrete GPUs do.ragenalien - Wednesday, March 9, 2022 - link
The first multi die GPU was released in 2021 by AMD. Are you saying the first prosumer GPU to use the tech? CDNA2 is in multiple AMD GPUs currently used in data centers.Ryan Smith - Wednesday, March 9, 2022 - link
CDNA2 places multiple dies on the package. But it does not try to make them behave as a single GPU. CDNA2 cards present themselves as two GPUs.ragenalien - Wednesday, March 9, 2022 - link
This is more due to software than hardware though. The interconnects for infinite fabric with the MI200 allow for up to 800 GB/sec between the two chips. AMD chose to have them present as two GPUs; Apple has chosen to do one. We have no benchmarks or data to show if apple's solution will be effective or not.techconc - Wednesday, March 9, 2022 - link
800GB/sec doesn’t compare well to 2.5TB/sec. Also, no, reporting as 1 GPU isn’t just a software thing.ragenalien - Wednesday, March 9, 2022 - link
It compares fine when it's just a GPU. That 2.5TB/sec also accounts for a CPU, Neural processor, and more, the GPU physically can't take over all of that bandwidth.Butterfish - Thursday, March 10, 2022 - link
Well, According to NVIDIA GPU totally can use that much bandwidth. CPU on the other hand don’t need a lot of bandwidth between two die.https://research.nvidia.com/sites/default/files/pu...
Ppietra - Friday, March 11, 2022 - link
I would imagine that most of that bandwidth would be used to access system level cache on the other die. Since the system level cache is accessible by CPU, GPU, Neural engine, etc, and communicates with the RAM controllers (both chips have RAM channels), I would think most of that bandwidth would be available to the GPU or CPU.mode_13h - Sunday, March 20, 2022 - link
> That 2.5TB/sec also accounts for a CPU, Neural processor, and more, the GPU physically can't take over all of that bandwidth.It's extremely unlikely to utilize all of those units to the max, all at once. That's a corner-case benchmarking scenario, not something that happens in the real world. In the real world, whatever you're doing might max out one thing or the other, but the app is going to be bottlenecked by aspect of the hardware, rather than have a perfect balance across the entire SoC's engines (if it even *uses* them all!).
What this means is that Apple basically just needs to design the SoC to support the most demanding aspect: CPU, GPU, or Neural (plus a little headroom, if they can manage it). Not all three, simultaneously. And if you did manage to max all 3, then I'm sure the SoC would do some significant clock-throttling.
mode_13h - Sunday, March 20, 2022 - link
> infinite fabric with the MI200 allow for up to 800 GB/sec between the two chips.Incorrect. There are 4 links between the dies of MI200. Each link can do 50 GB/s per direction. So, the aggregate is only 400 GB/sec. That's only a quarter of each die's memory bandwidth. So, if GPU cores are naively pulling data out of the companion's memory, that interconnect would be a crazy bottleneck.
That's probably why AMD *doesn't* treat them as a single GPU. They're merely presented as 2 GPUs that happen to share a package. MI200 seems primarily designed for use in systems that have more than one. And if you have a compute workload that scales to more than one MI200, then it probably doesn't hurt to treat the GCDs inside as separate processors.
Butterfish - Wednesday, March 9, 2022 - link
To be fair, AMD’s MI 200 is a lot more primitive when compared to this chips though. MI 200 only had up to 100 GB/s of bi-directional bandwidth between both GPUs which can be achieved with NVIDIA’s NVLINK without any multi chip packaging. This chips had 2.5TB/s which is very close to 3TB/s found by NVIDIA’s researcher to be the optimal inter die link bandwidth in their MCM research paper published in 2017, so Apple’s claimed that this multi die GPU will act as one monolithic GPUs may have some based in reality. If Apple actually pull that off, they will be the first one to crack the puzzle.ragenalien - Wednesday, March 9, 2022 - link
https://www.nextplatform.com/2021/11/09/the-aldeba...800 GB/sec bandwidth between the two chiplets.
Butterfish - Thursday, March 10, 2022 - link
Okay, my bad AMD’s MI 200 had 8 infinity fabric link between the two chips (8 X 100 GB/s)mode_13h - Sunday, March 20, 2022 - link
No, that's double-counting. Each of the two chips devotes 4 of its 8 links to its peer, but you only get 4 links' worth of bandwidth because they're connected to each other.For a system with only one MI200, you could get 800 GB/sec between them, by connecting up the external links between the two GPUs, as well.
mode_13h - Sunday, March 20, 2022 - link
> 3TB/s found by NVIDIA’s researcher to be the optimal inter die link> bandwidth in their MCM research paper published in 2017
That processor had like 16 dies. It's a lot easier to get big aggregate bandwidth numbers when you have lots of nodes.
In other words, it's a rubbish comparison.
Oxford Guy - Wednesday, March 9, 2022 - link
‘Apple is understandably excited about their accomplishment’‘its’, ‘the company’, or ‘the associated staff at Apple’. Each of those is appropriate but not ‘their’.
Corporations are not people, my friend.
Philotech - Wednesday, March 9, 2022 - link
> ‘Apple is understandably excited about their accomplishment’> ‘its’, ‘the company’, or ‘the associated staff at Apple’. Each of those is appropriate but not ‘their’.
> Corporations are not people, my friend.
That's a strange claim, especially considering your nick...
See here: https://www.grammarphobia.com/blog/2018/01/it-or-t...
In AE, companies are singular in *formal language*, but plural in more informal contexts. In BE, however, companies are much more likely referred to in plural also in formal texts.
mode_13h - Sunday, March 20, 2022 - link
> especially considering your nick...He could be the Oxford rubbish collector, for all we know. Of Oxford Mississippi. Nick would be just as accurate.
Oxford Guy - Wednesday, March 9, 2022 - link
‘After all, how would you even top a single 432mm2 chip that’s already pushing the limits of manufacturability on TSMC’s N5 process?’That’s a rather small limit considering how the 28nm Fiji consumer GPU was 596.
The Von Matrices - Thursday, March 10, 2022 - link
The Fiji chip came after nearly 4 years of 28nm products being on the market and all the process improvements achieved over that time. The N5 process is barely over a year old so the defect rate is probably still high enough for a 432mm2 chip to be considered as on the verge of manufacturability.BillBear - Thursday, March 10, 2022 - link
Yup. The 119 mm2 M1 die is tiny in comparison, but defects on N5 did start out high enough that Apple sells devices with a disabled GPU core at a discount instead of tossing chips with a defect in one of the GPU cores out.If you look at the Ultra, they are offering an option with 48 GPU cores enabled and another with all 64 GPU cores enabled, so wafer defects are still a thing on the two 420 mm2 M1 Max dies that are combined into an Ultra.
rmari - Wednesday, March 9, 2022 - link
At only 100 watts, the M1 Ultra should be able to be placed in a Laptop case.After all, gaming PCs use more electrical power.
blackcrayon - Friday, March 11, 2022 - link
Probably right, except Apple probably doesn't want to make a laptop that big & chunky (even the current 32 GPU M1 Max will throttle at extreme loads in a laptop the size of the current MacBook Pros).rmari - Wednesday, March 9, 2022 - link
I expect the new Mac Pro to have M2 Max chips with Ultrafusion fabric connector on TWO or THREE sides. This allows FOUR to EIGHT M2 Max chips to be connected into one monolithic processor!The GPU of the M2 will bring Ray Tracing to the Mac, finally. And the M2 Ultra will have between 128 to 256 GPU cores!
This would be a MONSTER chip as befits a Mac Pro.
usiname - Friday, March 11, 2022 - link
Now dont forget to replace your pantsrcfa - Wednesday, March 30, 2022 - link
More important would be real ECC memory.Vink - Thursday, March 10, 2022 - link
It's good, but how can I easily use it like an i5 + Radeon I have? At nothing !1.PC Gaming = NO
2.Audio: HiRes Player + Dac + Streamer + Amp = NO
3.Working: Autocad + SolidWorks etc etc ETC = NO
4.Working at Home: Team V., Any Desk, VPN, Remote D. etc = NO
5. Own Datacenter server accessible externally from anywhere with Remote D. = NO
6. Navigation anywhere without restrictions = NO
7. A very good protection available to be managed: Bitdef. Total Security etc etc = NO
8. 100% compatibility with the whole world: word, excel = NO
9. All Social Media very easily on my desktop: Whatapp, Instagram etc = NO
10. All movies and music very easily on my desktop: Netflix, Spotify = NO
... and of course a simple mouse with 2 buttons and scroll on my desk = NO atehnic magic mouse
AdvocateUK - Thursday, March 10, 2022 - link
Damn I didn't realise that you can't use a 2 button mouse with scroll wheel on a Mac! Oh, wait, you can. I could go on but I think I can sum your post up with just one more word....*sigh*
Vink - Friday, March 11, 2022 - link
I used a Logitech wireless mouse on an Air - a bad experience, then an ergonomic Microsoft wired mouse - another bad experience, then an Apple Mouse Magic - the worst experience (too small for my hand, inaccurate in sustained work ) ... plus a non-IPS screen - something very unpleasant overall ... I gave up. Otherwise: very fast, AL case super ok, excellent os (superb) ... no comments.*sorry*
Tomatotech - Friday, March 11, 2022 - link
So you bought the cheapest possible Apple laptop, then complained it’s not IPS? But at the time that fact would have been obvious to anyone who bothered to check what they were purchasing.Would you buy a PS5 then complain it doesn’t run Word, despite the maker not making that clear anywhere on their website?
In any case all Apple laptops now have IPS screens even the cheap models.
Logitech mice work fine with macs & so do all USB mice & keyboards. Maybe you struggled with the different mouse acceleration curves between Mac and Windows - that can be a big thing for some people (it is for me). Apple mouses have always been shitty, I don’t touch them. I’ll give you that one.
But that has no bearing on the list you gave. I only know half of the use-cases you mentioned but the half I do know about are bullshit.
Eg 7: ‘navigate anywhere without restrictions’ do you mean navigate around the OS? Yes you can turn off the restrictions any time you want or sudo / root your way in or even turn off sudo / account password entirely. Complaining about that ‘restriction’ is like complaining that Windows UAC illegally prevents you from doing whatever you want. It’s an ‘are you sure you know what you’re doing?’ step, but done rather better than UAC / Registry.
Focher - Thursday, March 10, 2022 - link
But the good news is you could still use it to post useless things on Internet forums. So, in that way, your life wouldn’t really change.blackcrayon - Friday, March 11, 2022 - link
Your list is somewhere between satire and "has almost no idea about Macs at ALL but enjoys writing about them in comment sections anyway". I'm going to lean toward the latter.rcfa - Wednesday, March 30, 2022 - link
🤣Of course you can, just about everything!
Just because someone doesn’t ship with the box or (temporarily still) runs in Rosetta2 doesn’t mean you can’t use or do it.
Jp7188 - Thursday, March 10, 2022 - link
I don't see multi-GPU as much of a hurdle for Apple. In what workloads are users going to be sensitive to GPU latency? Gaming is an area where people will freak out about less than perfect frametime consistency. Who is going to game on an M1 Ultra? AMD and NVIDIA have their work cut out for them to produce good multi GPU gaming chips.AdvocateUK - Thursday, March 10, 2022 - link
I agree. Gaming has just NEVER been a priority for Apple and I don't foresee that changing even with their recent moves a la Apple Arcade. The M1 Ultra may well be an awesome chip for gaming but I don't think we'll ever know.BillBear - Thursday, March 10, 2022 - link
Gaming on iOS is huge from a market share standpoint.Going forward, games on ARM based Macs will run on the exact same CPU and GPU cores using the same graphics framework that iOS games use. This means that targeting the Mac is no longer a niche with little market share. The same game code targets iOS and ARM Macs.
This worked well for XBox when it was breaking into a gaming console marketplace that Sony dominated, since XBox leveraged the same Direct 3D APIs, CPU cores and Nvidia GPUs that Windows games already worked with.
XBox had a tiny console marketshare compared to Playstation, but if you were already targeting Windows, it wasn't much more work to target XBox.
The same will be true of studios targeting iOS devices. Targeting ARM Macs is not much more work.
AdvocateUK - Thursday, March 10, 2022 - link
You’re right, iOS could well be (by numbers of devices) the largest gaming market in the world. How’s that affected Mac gaming? I don’t think by much, certainly not that I’ve noticed. Don’t get me wrong, I’d LOVE to see Mac gaming take off and was as excited as anyone about the possibilities when the M1 was announced. It just hasn’t happened. I hold out some hope for an Apple Silicon games console but let’s be honest, that’s probably not going to happen is it?BillBear - Thursday, March 10, 2022 - link
It hasn't effected Mac Gaming, because until now, coding Macs and coding iOS devices have been two different things.That is no longer true. ARM based Macs can play iOS games unchanged with zero effort on the part of a developer.
However, it makes sense to put in the minimal effort needed to at least add support for a game controllers (that support will work on iOS devices too) and support for Mac style menus.
So, sorry for your favorite narrative, but ARM Macs are going to be supported in a way older Macs never have been.
AdvocateUK - Friday, March 11, 2022 - link
My favourite narrative?I’m a Mac user and have been for over 25 years. Our household is a complete Apple household from the 27 inch iMac in the office, the M1 MBP I take to work, the two AppleTV’s, 12.9 inch iPad Pro, in fact multiple iPads Pro, multiple iPhones, HomePod speaker, original Apple HiFi speaker, iPod Classic (somewhere), and the many old and not used anymore Macs I have lying around.
This isn’t a narrative it’s a fact, despite everything Apple has done developer’s just aren’t developing games for the Mac in the numbers that You, I, or anyone interested in gaming would like them to.
I agree that the AS Macs are likely to be better supported by Devs but so far they haven’t been the panacea we expected.
techconc - Friday, March 11, 2022 - link
Yeah, it's unfortunate that Apple just has no real interest in the gaming market. Sadly, they think crappy indie games on Apple Arcade checks that box. I suppose it does for the Candy Crush crowd, but not for actual gamers. I can't help but look at the Apple TV as such a missed opportunity for them.As I contemplate purchasing a Mac Studio, I have to wonder whether the M1 Max is sufficient enough. For video / photo editing, the M1 Max is more than enough. If only there were some really high end games where the Ultra would make a real difference...
BillBear - Friday, March 11, 2022 - link
I should have said "a very popular narrative".I am reminded of the people who stuck to the narrative that AMD couldn't possibly compete with Intel back when they hired Jim Keller and put him back in charge.
We are now moving into an era where Mac gaming is changing from a not terribly profitable niche to Mac gaming piggybacking off of one of the most profitable areas in gaming.
Cheap ports to a ARM Macs changes the economics of game development.
AdvocateUK - Saturday, March 12, 2022 - link
Point taken, thank you.I’ve been around long enough now that I know anything is possible. From the original days when nobody believed AMD could compete with Intel - and then came Athlon and 3dNow, then when AMD fell behind again that was supposedly their end, they were doomed. Same with Apple, I bought into the Apple ecosystem around the start of OSX (my first laptop came pre installed with both OS9 and OSX) when everyone was saying it was all over for Apple, and never bought another PC since.
But I’ve also seen Apple promise time and again that “this time we’re serious about games” and rolling out the likes of Jon Carmack to prove it only to see Max games withering as always.
I agree that this time they have a real chance at it, but to do so they need to be pushing a box like an AppleTV Pro with the latest Apple Silicon in it and a crap load of onboard storage - and keep it updated, every single year.
But they haven’t, and I worry that game devs have been burnt once too often to commit serious resources to developing AAA games on Mac.
Jp7188 - Friday, March 11, 2022 - link
I don't know about "zero" effort. Many iOS games would lose a lot going from a small touchscreen to a large non-touch keyboard/mouse interface.But more to the point, what iOS game is going to benefit from the power of an M1 Ultra? What person would buy an M1 to take their iOS gaming to the next level? Now ask those same questions about NVIDIA and AMD GPUs. Gaming is the primary driver for PC GPU innovation and barely a thought for Apple M1 GPU development.
Game frametime consistency and latency are the main reasons we don't have multi-GPU gaming today; it's one of the reasons SLI died. The intended workloads for M1 Ultra don't have those concerns.
BillBear - Friday, March 11, 2022 - link
>Many iOS games would lose a lot going from a small touchscreen to a large non-touch keyboard/mouse interface.Yes. Just like PC Games based on a mouse and keyboard didn't work without some effort when moved to an XBox. You had to add support for a gaming controller.
Adding support for gaming controllers on an iOS game means those gaming controllers (XBox or Playstation) will also be supported on an ARM Mac.
As I said, there is not zero effort required to make a good Mac game, but it is definitely minimal effort.
Not to mention gaming engine makers who can't support iOS without also supporting the Mac now.
andrewaggb - Monday, March 14, 2022 - link
I'm curious about their multi-gpu for gaming as well - I suspect it hasn't solved the hard problems. It's not like it's a unique or new idea, I had Voodoo2 SLI like everybody else back in the day. But that technique didn't scale to modern games. ATI had a rage fury maxx that was a dual chip AFR render (which added latency) and there was the nvidia SLI of the past decade, and various other AMD implementations which were often game dependent. Both NVidia and AMD have said it's not as simple as putting two chips on the same card. If it were a 3070 would probably be 2 3060 chips and a 3080 would be 3 or whatever.maloxplay - Friday, March 11, 2022 - link
#applesempretroppoavantiApple manages to overcome each time, this sector is very complicated and full of pitfalls but thanks to the determination of the Company, they have also managed to become successful in the world of chips and thanks to this new way of connecting the 2 dies they have really raised the rod. the Apple Silicon product was a revolution and also the new ultrafusion system will be a reference for the entire industry.
johanpm - Friday, March 11, 2022 - link
Great performance, I'll give them that. But running 128GB of RAM without ECC is scary. Maybe the graphic/video people don't really care about a bitflip once in awhile.mode_13h - Sunday, March 20, 2022 - link
As it uses LPDDR5, shouldn't it at least benefit from on-die ECC?rcfa - Wednesday, March 30, 2022 - link
on-die ECC is, to be nasty, a means to increase chip yields, not to increase reliability.That said, the key thing of ECC is that the OS knows it has verified data, and for that ECC has to operate during RW ops, not during refresh, and the RAM needs to be able to signal failures to the OS, not silently fail.
Especially when running modern file systems this is key, because it’s easy to corrupt large swaths of storage by writing corrupted fs metadata
mode_13h - Thursday, March 31, 2022 - link
> on-die ECC is, to be nasty, a means to increase chip yieldsOverall, yes. I get that. However, it's conceivable that Apple might've contracted with a DRAM supplier to allocate more on-die ECC bits so they can have a net-increase in reliability, no?
> the RAM needs to be able to signal failures to the OS, not silently fail.
Depends. If you have replaceable RAM, then it's valuable to know that you have a DIMM that's beginning to fail.
The other reason is so the OS can know a certain address range has a defective memory cell and it can reserve that page so it's taken out of circulation. However, this *could* also be handled on-die, where the DRAM itself could remap bad rows. It'd be far better for the OS to do it, of course, so that the DIMM didn't have to implement physical address translation.
mode_13h - Thursday, March 31, 2022 - link
> Especially when running modern file systems this is key, because> it’s easy to corrupt large swaths of storage by writing corrupted fs metadata
That's why I'm a die-hard fan of FAT filesystem!
LOL, just kidding. No, seriously: it's why I would only run a fileserver on hardware with ECC RAM.
Tomatotech - Friday, March 11, 2022 - link
I can’t see how AnandTech can claim this is the ‘last member of the M1 family’. The Apple Silicon transition is not yet complete. We still have the Mac Pro to go.That could be the M2 or it could be yet another version of the M1.
M1 Ultra Plaid.
andynormancx - Sunday, March 13, 2022 - link
Apple very explicitly said in the presentation that the Ultra was the last member of the M1 family of chips. So AnandTech aren't claiming it, that is what Apple said.Tomatotech - Monday, March 14, 2022 - link
Thanks for the clarification. That phrase 'We're adding one last chip to the M1 family' wasn't in the AnandTech liveblog, but Ars Technica reported it in their liveblog.That's interesting, because speculation is that the Mac Pro will have 2x M1 Ultras glued together - that lines up with all the previous rumours about the new Mac Pro's upcoming capabilities. But every time Apple have stuck two chips together, they've given it a new name. M1 Ultra = 2x M1 Max = 4x M1 Pro.
So either this is NOT the final (and apple didn't use the word 'final') M1 chip, OR it will be called the M2 (unlikely) OR it will be 2x M1 Ultra chips in separate packages, not connected together.
Using 2 separate packages seems un-Apple like, and introduces unnecessary complexity / issues. So my money is on the first option - that Apple will introduce yet another M1 variant for the Mac Pro. "We said in March that we were adding one last chip, the M1 Ultra. Well, guess what, we've made a new one! Witness the M1 <Plaid>!"
Historical evidence:
(1) Apple constantly one-upping themselves each year.
(2) M1 Max being inferior (!) to the later release of the M1 Ultra.
michael2k - Tuesday, March 15, 2022 - link
Connecting two M1 Ultra is no more complex than connecting two GPUs in the current Mac Pro. At the worst they use PCIe5 x16 for 128GB/s between chips and at best they use UCIe and if they can keep the two chips 2mm apart they can connect them at 1.2TB/sandynormancx - Tuesday, March 15, 2022 - link
It seems far more likely to me, as already discussed here, that they’ll wait for the M2 for the Pro and that the M2 will have extra interconnects to allow them to combine more than 2 dies at a time.That might be a struggle within their original migration target, but it is entirely possible that they might pre announce the new Pro sometime before it is available, giving them a bit more time.
involuntarheely - Tuesday, March 15, 2022 - link
that's very unlikely because Apple will never give up scale economies for M2. they obviously went this path with the M1. they started with iphone and by the time the A13 (or whatever the number was) became M1 they had produced millions and brought production costs down. then they came up with M1pro/max when those became cheap to make. and so on.they will introduce the M2 on Air/mini, then MBP 14/16, then iMac, then Studio, then Pro, which I'm guessing follows the sales numbers of the respective devices (save for the mini, I guess).
there would be a huge cost in starting M2 sales with the Pro.
BillBear - Wednesday, March 16, 2022 - link
It's also possible that the chips in the Mac Pro line will be considered a different chip family altogether.One that focuses on throughput without caring about power draw or production costs.
I wouldn't be surprised to see massive amounts of HBM,for example, that would not be cost effective in a consumer product like an iPad, but is already in use in high end products from others.
Bevans868 - Saturday, March 12, 2022 - link
Is it possible that synthetic benchmarks are currently capable of accurately Apple SOC chip architecture? When you add encoders/decoders, unified memory, neural engines I think they underestimate real workloads.brucethemoose - Tuesday, March 15, 2022 - link
Well that goes both ways if you, for instance, start talking about nvenc and CUDA. Or the ASICs in the AMD consoles.The M1 series is very good at media processing though, I don'f think anyone will dispute that.
Mehen - Monday, March 14, 2022 - link
Either Apple does not know how to label a graph properly, or I think there are some major misinterpretations going on here.The GPU performance vs. power graph.
Y-Axis = "Relative Performance."
Not "Performance." It says "Relative Performance."
Relative to what? It's clearly not relative to the 3090 setup (otherwise the 3090 results don't make sense, it would have to be a straight horizontal line). From how I read it, "Relative Performance" is "Performance / Watt." Let's call it P/W. You can think of P as Performance Points, kind of like a 3D mark score or something.
Let's look at the maximum performance on the graph. The top of the M1 Ultra looks to hit about 108W (just a little below the midpoint between 100 and 120 on the X-Axis).
So at 108W, the M1 Ultra hits ~180 on Relative Performance (P/W).
So 180P/W x 108W = 19,440 Points.
The highest point of the 3090 setup is about 315W. The way the graph reads, I'd peg this around 170P/W, but lets use the 90% figure they also used on the CPU graph, and give the Apple chip even more of a lead, (180P/W x .90) and we get 162P/W for the 3090 setup.
So at 315W, the 3090 set up is ~162 on Relative Performance (P/W).
So 162P/W x 315W = 51,030 Points. (It would be even more if I used 170P/W)
So the M1 Ultra is still very impressive, scoring about 38% of the 3090 setup (19,440 / 51,030), at only 34% of the power (108W / 315W).
Of course, if they actually provided some sort of units or benchmarks for the Y-Axis, they could clarify this whole mess of a graph. If it really is that fast, label is properly!
Don't these numbers seem much more reasonable? And it seems to fit better with how previous chips compare with low TDP versions of the graphics cards (mobile 3080's for example).
Thoughts? I feel like Apple is intentionally misleading us with that "200W less power" notation.
bifford - Monday, March 14, 2022 - link
"Relative performance" does not mean performance per watt. "Relative performance" just means that the ratios between the two chips are accurate but it does not have a specific unit. This is very common nomenclature in science when comparing two processes. It is probably based on a summary of their internal benchmarking process scaled to make the numbers pretty.Mehen - Monday, March 14, 2022 - link
Gotcha. The scaling thing makes sense, I think that is what I was missing. And... well... I guess my next gaming rig is going to be a Mac then!Tomatotech - Monday, March 21, 2022 - link
They're lovely chips but I really really wouldn't bother buying a Mac Studio for gaming use. You need at least $4000 to get an M1 Ultra, it won't run Windows well (yet) and there is no guarantee that any Mac games will be properly coded to make use of the M1 Ultra power.Better to, if you can, pay for gigabit internet service (£25 a month for me in the UK with 2 ms ping) plus a subs to GeForce Now ($20/month or $200 per year) for 3080 tier gaming on your Mac / PC
rmari - Tuesday, March 15, 2022 - link
I would wait for the M2 Max Extreme with a 48 core CPU and 128 core GPU for the Mac Pro.That would be a blast.
quiet-cheese - Saturday, March 19, 2022 - link
“While the details very from implementation to implementation”probably mean to say vary instead of very
jalandhara - Sunday, March 20, 2022 - link
And some Core i5 chips are faster than some Core i7 chips.for more https://gbapk.in/fouad-whatsapp-apk-download/
mode_13h - Monday, March 21, 2022 - link
spammer.valuearb - Sunday, March 20, 2022 - link
Since it's a well know secret that the new Mac Pro is coming with a 40 core CPU, how does this interconnect work with four dies?mode_13h - Monday, March 21, 2022 - link
I wouldn't bet on it. name99 said the interrupt controller basically can't scale further.Plus, the Pro needs to handle more than 256 GB of RAM, to be a proper replacement for the prior generation.
Finally, the cores in the M1 are at least a generation old. So, that's all the more reason Apple will probably use a new die.
mode_13h - Monday, March 21, 2022 - link
I could actually see Apple taking someone else's Neoverse V1-based server CPU, for use in the new Mac Pro. Maybe they'll put an M1 Ultra on a PCIe card and use it as a dGPU?scottrichardson - Thursday, March 24, 2022 - link
A YouTuber did some rather sound research, including paying for paywalled articles from TSMC which outline the interposer tech used in the Ultra, and an additional version of the interposer that supports four-way linkage, plus a data backbone that will be used to connect to slower large scale RAM banks. Likely based on M2, so 48 cores (8 extra efficiency cores), 160 GPU cores, and up to 256GB unified memory with access to a further 1TB of more classical RAM. Also stated that the same backbone will be able to access additional GPU cards, and suggested that Apple may sell their own GPU PCIe cards like they did with afterburner.mode_13h - Friday, March 25, 2022 - link
> including paying for paywalled articles from TSMC which outline ...> an additional version of the interposer
If, at this point, your Spidey senses don't tingle, they must be broken. There's no way TSMC is going to dish sensitive details about their Apple-specific solutions. And there should be no doubt that this interposer is *very* Apple-specific. But nobody is going to fact check it, especially if the sources are supposedly pay-walled.
Unless a leaker has a strong track-record, I pay them no mind. It's too easy to dress up mere speculation and the incentives are tilted strongly in favor of the sensational, with almost no negative consequences for being wrong.
Qasar - Friday, March 25, 2022 - link
" A YouTuber did some rather sound research " which one, post it here so we can all go and see it.sounds like pure BS to me, which is what it looks like Mode also implied.
xol - Monday, March 21, 2022 - link
Compare the old article : https://www.anandtech.com/show/17024/apple-m1-max-...quote "Apple’s die shot of the M1 Max was a bit weird initially in that we weren’t sure if it actually represents physical reality – especially on the bottom part of the chip we had noted that there appears to be a doubled up NPU – something Apple doesn’t officially disclose"
It looks like the old die shots were deliberately misleading - it's easy to compare the old and new images. Where the "2nd NPU" was appears to be where the 'interposer hardware" is today.
Probably intentional on apple's part but the old article needs and update
mixmaxmix - Monday, March 21, 2022 - link
s22 ultra's review when upload??mode_13h - Tuesday, March 22, 2022 - link
Given that they lost Andrei and Ian, it might be a while (if ever).Enricosba - Tuesday, April 19, 2022 - link
Do you know where they are writing now?Thanks
Enricosba - Tuesday, April 19, 2022 - link
Do you know where they are writing now?Thanks
kath1mack - Thursday, April 14, 2022 - link
InterestingKatety - Saturday, May 14, 2022 - link
Thank you everyone for sharing the useful information that I have been looking for for a long time, <a href="https://runaway3d.com/">run 3</a>sunailkarim - Wednesday, May 18, 2022 - link
wow great article <a href="https://apktami.com/apk/rts-tv-apk-latest-68/"... tv apk</a>