It's a bit disappointing that Intel seems to have gotten it's shit together. I would have preferred to see them suffer for a little while longer, so that AMD can grow enough that we can be sure that they'll survive Intel's eventual rebound.
But I guess it's still better than having Intel be burned so badly that the investors chose to sell off their foundry business, or stop making consumer CPU's, or something else along those lines. It's already a shame that everyone is talking about how they'll abandon Optane R&D.
In the end, we need someone else to make sure TSMC behaves, and Samsung seems to always be two steps behind them.
Definitely welcome to see the competition light up. It can only be good for the end users to have more suppliers and options. I am a little skeptical that Intel will be able to pull off 4 processes on EUV in 2 years, but it will be impressive if they do.
Samsung & TSMC have had years of experience with EUV but both still seem to have some issues with their latest processes.
"Definitely welcome to see the competition light up. It can only be good for the end users to have more suppliers and options." Unfortunately, the presence of Intel in the market usually brings fewer suppliers and options - see Intel's past actions, some which generated fines and some which were settled out-of-court.
AMD had control of the consumer market for a scant 2 years before jacking up prices to the moon. Intel's near monopoly from 2007-2016 prices remained mostly stable at roughly $300-350 for high end consumer chips.
Granted, they were quad cores, but still, a monopoly under either company is horrible. I dont want AMD getting any lazier then they already have.
so 50 buck increase on MSRP is considered " jacking up prices to the moon. " yea ok sure. IF you are referring to the prices in the store, they werent all that bad where i am, picked up my 5900X for $699 cdn i think it was last year
intels prices went up a lot more then amds here before zen was released, and then the top end intel cpu dropped $1k shortly after that release.
between intel and amd, amd so far, as been the lesser evil of the 2.
You have always been able to get reasonably capable AMD CPU's with Zen architecture and integrated graphics as low as $60 (Athlon 3000G) for years now, while powerful entry-level CPU's like the Ryzen 3200G had been selling around $120 prior to the pandemic (you'd have to go for something without integrated graphics to get in that price range now)
This is all on par with Intel at the low-end. However, at the low-end, Intel is faster. AMD isn't competitive until you spend $150-$200. That really comes down to two things: for AMD to turn a profit with their smaller volume, they must sell more expensive chips, and the obvious: AMD is dependent on TSMC, who dictates supply, and to some extent, cost.
Sonic’s menu includes the most traditional aspects of fast-food cuisine: hot dogs, chilli dogs, onion rings, soft drinks, slushies, milkshakes, hamburgers, french fries, corn dogs, sundaes and banana splits.
They have continuously updated their fast food menu to satisfy their customers’ requirements.
The fab business being spun off is something some investors were already calling for so I think it's fair to say this is far more good than bad.
AMD seems to be in a pretty solid position, even against SPR, and they've been making a lot of big moves so as long as none of it blows up I think they're good even with the imminent Intel rebound. It seems like each company is carving out their own priorities in the high profit sectors, and given production limitations (see TR Pro 5xxx supply issues) I think it'll be a few years before they're stomping on each other's toes again.
Yes, but 3D cache puts a big dent in HBM's advantage (for some workloads).
Has Intel even announced the programming model supported by SPR-HBM? Other than using it as L4 cache, I think it's not going to see a lot of use for the first year or so, while software support begins to ramp up for it. I see SPR-HBM as a software development vehicle for hierarchical memory, more than anything else. Sure, there will be some niche applications that benefit quite nicely from it.
I don't believe they have announced the programming model, and the first chunk of them are going straight to HPC from what I've been able to find out. I'd imagine these will be the last SPR models to hit the market if they do at all.
the hbm is probably on there for AVX512/AMX throughput in things like machine learning, not really displaced by milan-x. it'll end up competing with genoa.
Seems is the operative word. I'm sure Intel will have a node CALLED 18A. And they will ship something on it in 2023Q3. But will it actually be a significant change from the earlier node? And will the item shipped be of any relevance?
intel Arc is an interesting test case for the supposedly new, supposedly honest Intel. That was going to ship Q1 2022. And, supposedly, in theory, it shipped, possibly, somewhere in the world, on March 30 which I guess is technically Q1 2022. But can you actually buy one? And if you could, would you want to?
Now apply that same outcome to this process roadmap/fantasy...
Whatever it may be CALLED, it will be competitive with a similar TSMC offering. The whole reason they renamed things in the first place was because they were the only ones adhering to such tight standards. 'Intel 7' is quite similar to TSMC N7, and Intel 4 is expected to be better than the TSMC '4nm' equivalent.
Now that they are getting their process issues straightened out, they need to improve their Core design to address power efficiency concerns.200+W for a CPU is NOT cool.
They renamed it as a marketing exercise. If Intel's 10nm Enhanced SuperFin (it’s also been called 10nm+++ and 10nm++) was a bear to communicate.
Also if it were competitive with TSMC N7 there wouldn't be a discussion about their competitiveness. Intel 7 has been in production since fall 2020; Apple was using TSMC N7 in fall 2018, two years earlier.
Now if you're saying Intel has finally caught up to N7P and that later this year they will catch up to N5, and finally be current with TSMC's N4 by next year... you're guessing as much as anyone else.
> they will ship something on it in 2023Q3. But will it actually be a significant > change from the earlier node? And will the item shipped be of any relevance?
Yes, key questions. Many of us remember how they seemed to ship a few trays of Cannon Lake, just so they could tell investors they delivered on some revision of their 10 nm timeline.
> intel Arc ... That was going to ship Q1 2022.
Yes, March 30th was suspiciously close to the end of the quarter. Funny enough, "Intel shill" site Anandtech didn't even cover their launch announcement, unlike many other tech sites.
> Now apply that same outcome to this process roadmap/fantasy...
Comparing manufacturing process to a graphics product is not really apples-to-apples.
For me, the elephant in the room is still Intel's 10 nm debacle. That's reason enough to doubt everything Intel says about their manufacturing tech.
They may have been more honest about 14nm, 14+, 14++ but their schedules and roadmaps were pure fantasy for years (going back to the transition to 14nm). And this article is all about the "new improved marketing" of the schedule.
Intel responded to the Optane rumors. Gen 3 optane is still on its way. They also added optane support to the new agilex-M fpga. They also indicated plans to create a cxl controller that handles optane.
Gen 3 Optane will unfortunately be using the Gen 2 3D XPoint media, but with new control options. I too cross my fingers that there will be more generations forthcoming with more layers of cells as I know that's been a primary cost factor. There's a lot going for the technology, but unfortunately a lot of barriers for another client segment, but I'm still holding out hope.
Oh, I dunno, AMD was already getting to the point where they needed a good competitor again. Look at how long it took them to finish the Zen 3 product stack! It wasn't until Alder Lake that they took low to mid range parts seriously again. Not to mention how Zen 3 will probably get to celebrate it's second birthday before Zen 4 really kicks off.
> AMD was already getting to the point where they needed a good competitor again.
I doubt it. I'm sure zero people at AMD were kicking back and snoozing before Alder Lake.
It's hard to appreciate how much work it is to scale up your customer base, but that's exactly what AMD has been doing, the past few years. Each one of those customers takes internal resources, wants special features, finds new bugs, etc. Rapid growth of a business is actually hard to do well, especially without compromising your execution.
> Zen 3 will probably get to celebrate it's second birthday before Zen 4 really kicks off.
Lots of things have to align, for a successful launch. Remember how Zen 3 used the same manufacturing process as the Zen 2 XT processors and the same platform. Zen 4 is on a new process node and new socket.
Also, try to remember that Zen+ was almost a direct port of Zen 1 to 12 nm. With that in mind, the gap between Zen 1 and 2 was also pretty big.
So...did you just sleep through all of 2020 and 2021? Did you miss AMD jacking up their 6 core's effective price by 90% for a 20% uplift? The whole "no zen 3 below $270" thing that lasted until alder lake launched, then suddenly BAM price drop BAM price drop BAM new products?
Seems pretty obvious to me that AMD's old attitude of falling asleep at the wheel the moment they get the lead, which destroyed them with the first gen phenom, the HD 6000 series, and GCN 1.2, is alive and well.
" So...did you just sleep through all of 2020 and 2021? " did you ? or did you not see what the world went though, and in still kind of going through? you DO realize we have and are still dealing with Covid-19, right ?
> then suddenly BAM price drop BAM price drop BAM new products?
When Intel launches something that's actually competitive, the logical expectation would be for AMD's demand to slacken. And that lets them justify re-pricing and adding lower product tiers.
Can you imagine what would've happened if AMD did that a couple years ago? The products would've still sold out and most end consumers would've still paid similar prices, but the difference would've just gone to retailers and scalpers.
> AMD's old attitude of falling asleep at the wheel the moment they get the lead
Zen 4 will tell us if they took their foot off the gas. Until then, it'll be interesting to see how the 5800X3D benches in the real world. Even if it comes with heavy tradeoffs, I look at that as a learning experience, for them. I'm sure the 3D chiplets in following generations will benefit from the experience.
> first gen phenom, the HD 6000 series, and GCN 1.2, is alive and well.
In some cases, AMD stumbled into financial problems and had to make painful layoffs. In other cases, the competition just came back stronger. I'm not saying they didn't also have an under-performing mindset, which you can sort of see from the interviews with Jim Keller.
How many years we got skylake refresh? AMD still make good improvement, not every year, but at least on 2, and Zen4 is expected to delivery 25-30% ipc imrpovement.
Yeah they were making 20-25% gains per gen and the moment they passed intel the prices started going up, in the 6 core's case, 90% in a single gen for that 20% uplift.
Never forget, the 3600 was a $169 product replaced with a $299 product. If you were a budget builder, go pound sand and buy intel.
> the 3600 was a $169 product replaced with a $299 product.
No, the 3600 launched at $199. It wasn't replaced until the 5600, which recently launched at... wait for it... $199!
The 3600X launched at $249 and the 5600X launched at $299.
As @Qasar, points out, that launched in Nov 2020, when demand for CPUs was abnormally high and Intel was still pushing Skylake cores on 14 nm. So, the $50 price increase was understandable, if unpalatable.
You're making at least two errors, here. You're equating 6-core with 6-core, regardless of positioning, and you're comparing the discounted *old* model against pricing of the new model. 3600 didn't *launch* at $169.
We could just blame supply problems & demand spikes on AMD and pat ourselves on the back for our moral outrage, but if you want more insight into *why* the prices shifted, then you can't ignore how the context changed.
The Zen 3 product stack has been waiting for demand to slacken. AMD could have released these chips sooner, but they'd have only been cutting off profits for themselves.
We don't have to /like/ it as buyers, but it's a cold calculus, and if we want AMD to survive as a competitor then some number of decisions like this will have to be made.
> The Zen 3 product stack has been waiting for demand to slacken.
I'm not sure about that. Zen 4 needs TSMC's N5 to be sufficiently mature (not to mention capacity) and DDR5 to have good availability. AMD should've wanted to launch it in Q1 of this year, especially if you consider all the discounting they're doing on the Ryzen 5000 product stack.
> AMD could have released these chips sooner
So, I reject the assertion that AMD simply didn't *want* to launch Zen 4 any sooner than they are. I think they're getting it out the door as soon as they reasonably could.
>The Zen 3 product stack has been waiting for demand to slacken. AMD could have released these chips sooner, but they'd have only been cutting off profits for themselves.
Isnt it funny that it just so happens the demand slackens as SOON as intel's alderlake goes on sale? Almost like that's a massive cope.
> Isnt it funny that it just so happens the demand slackens as SOON as intel's alderlake
No, it's perfectly logical. Alder lake is much more competitive against Zen 3 than either generation before it. Of course that's going to steal some of Zen 3's demand! We'd be fools to pretend otherwise.
You're so concerned with competition among CPU makers, but what about ASML? In particular, their exclusive partnership with Intel on high-NA (as in aperture -- not sodium!) is a little troubling.
> it's still better than having Intel be burned so badly that > the investors chose to sell off their foundry business
I still think that would be for the best. I'm glad Intel is starting to treat it more as a separate business, since that will ideally be the eventual outcome.
"18A development has been moving so well that the company’s R&D operations are now on or ahead of all of their development milestones" That does not mean that they will reach risk production as expected, or "real" production numbers as expected. It means that some hard problems were solved, and some other hard problems remain. As for the additional space needed... improved processes usually take longer (or much longer) per wafer, and (due to the increase in transistor counts) the number of processors per wafer didn't increase much.
AMD has shown that when they get on top- they are just as bad as Intel. Look at their latest offerings= r5 5500 4500 and radeon 6500, all total trash products.
So, you think the determining factor was AMD's recent dominance? I say it's the historic chip shortage.
The RX 6500 was designed to be cheap and crypto-unfriendly by cutting die size and memory. The Ryzen 5500 and 4500 are probably just selling off some low-binned APUs at a discount. Right now, I think it's difficult for AMD to provide a better answer to Intel, at the entry level.
"all total trash products" Not really, though. You picked their weakest and cheapest products, all of which serve specific niches and all of which undercut their competition (dramatically in the case of the 6500 XT).
Yes indeed in the end we all need someone else to make sure TSMC and likewise offshore friends are kept in check! With Taiwan now being seriously in the crosshairs by China and global superpower territorial expansionism is here to stay, I wonder what that will bring next to our own tech-doorstep? Taiwan as we speak is also not welcomed being a member of the U.N or even considered by many to be a sovereign nation. The good news however is that the USA is now being seen by many as a ‘Safe Harbor’ country. Intel here in Oregon across four campuses in Hillsboro – just ten miles west of Portland just opened its $3-billion expansion with close to 21,000 employees. To date over 65,000 job applications have been received and with people vying for the benefits such as 401K, unrestricted medical and dental, paid vacations, onsite chld care, guaranteed long-term fulltime employment and the top salaries Intel offers. Most certainly and ignoring the technical aspects discussed here for a minute all of this makes me a fan of Intel!
Intel will abandon Optane once they can no longer pretend it is dead.
If it wasn't fully dead, Micron wouldn't have simply thrown their share of it away. This is rather sad as had it worked, it wouldn't have been a stretch to make huge DRAM caches for vast Optane "main storage". But it don't expect it to ever be more than a really expensive alternative to flash marketed only to enterprise customers.
Well it was expected that R&D path that Gelsinger took Intel which it needs is showing. Good to see more R&D on the Lithography side. Especially when we have the total domination by TSMC. A balance needs to be there. Note Samsung is also in the race of GAAFET 3nm same as Intel 3 or 20A / TSMC 3 or 2.
But unfortunately, all this tech is going to be wasted on the garbage BGA trash more than HEDT or Mainstream Desktop because that's where these investor riddled companies are at. And worse is E-Core BS that Intel created because of their Intel Core uArch being too old, Nehalem. So until 2025 if the Z990 or Z1000 series chipsets, the Intel Core branding will relegate itself to the pathetic ARM copy cat design of growing more E Cores trash. Maybe by 18A+ with NAEUV Intel will redesign their x86 with a successor to Core. Also hopefully no more BS of awful security holes and tons of papers from Technion students, I have a benefit of doubt how Intel Haifa behaves given the history with NSO Group and Pegasus feud with Apple corporation.
Plus their Xeon will give competition to Bergamo after a whole year, by 2023 Bergamo will be out, and 2024 is when Xeon with E Cores only will come. So Intel is decided on this big small path unfortunately so is AMD but AMD so far doesn't incorporate both into one CPU. I hope they do not do that for Zen 4 or any CPU on AM5. My guess is AMD will make 2 SKU classes, absolute top will be high performance Big trad x86 and the 2nd class would be c cut down versions.
Finally what's the use when the PC itself is seeing weird changes such as Windows 11 going more for BGA Touchscreen dumpster rather than poweruser based UX like Windows 7. And by 2025 and up, not sure how GaaS and these pathetic Service models will fare. So all this tech and growth wasted on such is a massive disappointment.
Intel will soon have 16 e-cores on the desktop chips ... Raptor Lake... and the rumors are that will increase to 24 e-cores on the chip after Meteor Lake.
I see wccftech reporting a 56 core alder lake-x chip. Will that be mostly e-cores?
"Finally what's the use when the PC itself is seeing weird changes such as Windows 11 going more for BGA Touchscreen dumpster rather than poweruser based UX like Windows 7."
there was a time, some years ago, when some wag (no longer remember who; might have been Dvorak) allowed as how there were three killer apps for the PC - spreadsheets - word processing - spreadsheets
today we might add a fourth - email
either way, a 486 with a decent bit of wifi will do for most folks.
Web, online shopping, and social media are obviously killer apps. Sure, you can do them on a phone, but they're all better on a PC.
And that's leaving out the granddaddy of them all, which is gaming. There are gaming and education apps that don't port well to consoles or phones.
> a 486 with a decent bit of wifi will do for most folks.
No. The web is a moving target. What CPU power it took to brows the web of the 1990's over dialup (or a 1 mbps cable modem, for the lucky) will no longer cut it. A Raspberry Pi is lightyears faster than any 486, and it's still painful for certain things.
I think this little troll post backfired on you. It seems as if you remember nothing of the past 2 decades.
486 is theoretically limited to 4 GiB of addressable memory. However, you would be hard pressed to find a mainboard supporting 64MiB, with 32 MiB a more credible alternative. So, while the 486 could be fast enough (which I doubt), it will most certainly not have enough memory.
I had a Pentium Pro with 128 MB of RAM, but I forget whether it was FPM or EDO. I ran a few of my own benchmarks on it, and I think it managed about 90 MB/sec for read/write. On any 486, it should've been some fraction of that. So, even if you could pack it with say 2 GB of RAM, actually using it would've been another thing entirely.
I think the first machine I had with 2 GB was a Pentium 4, which ran at 3.2 GHz compared to the 33 MHz of a typical 486.
Once AMD goes big.Little in 2 years or so, and every CPU manufacturer, from Intel, AMD, Qualcomm, Mediatek, and Apple are all using big.little, will people who have no formal training or real world experience in CPU design still be calling the entire industry of competing companies and engineers 'idiots'?
And you're mad Windows 11 accommodates touch? I like using a touch screen convertible device.
He may be unbalanced, but he's not that wrong. It remains unclear that E-core is a great match for the desktop.
E-core clearly has massive value going downward (iPhone, iPad, laptop; but also Apple Watch and HomePod mini). And E-core probably has a useful role in the data center for servers whose primary job is to move data around with minimal manipulation thereof (things like memcached servers).
But on the desktop? Look at how Apple dialed BACK the number of E-cores for M1 Pro and Max. It's surely significant that the company with the most experience of these things made that decision?
Now there is a possible direction E-cores could go, which is to provide a sea of generic throughput computing. This is essentially the bet that Tenstorrent (Jim Keller's new crowd) are making. And you could imagine such a sea of generic throughput being useful for NPU, GPU and ISP type applications. But that's a hell of a bet and it's unclear that it will pan out; Intel made a similar bet with the first version of Larrabee as GPU, and, well, we all know how that went -- it's hard for generic compute to compete against specialized HW once the problem domain is well-enough specified...
So for Intel to launch, say 8P- + 24E-cores is a hell of a bet that enough important SW companies figure out, in just a year or two, a way to extract value from a flock of chickens. The industry, more or less, has been working on this problem since at least Transputers and the BBN Butterfly, with limited success, so it seems unlikely that we'll see the relevant breakthrough soon.
> Look at how Apple dialed BACK the number of E-cores for M1 Pro and Max.
Apples and Oranges, I say. Apple's P-cores are a lot more efficient than Intel's, so Apple probably doesn't have to worry about thermal throttling with the number they were willing to pack onto the Pro and Max dies. And it's not like they have so many P-cores, either. Where E-cores really come into their own is when you start to scale to even higher core counts.
> And E-core probably has a useful role in the data center for servers whose primary > job is to move data around with minimal manipulation thereof
Intel's E-cores aren't *that* weak. Gracemont is about 65% as fast at integer workloads as Golden Cove (both, single-thread), in Alder Lake. Once you scale up core counts, the P-cores would have to run at lower clocks, thereby narrowing the gap further.
Also, Sapphire Rapids will introduce special-purpose "DSA" engines for simple data movement tasks.
> a possible direction E-cores could go, ... a sea of generic throughput computing. > This is essentially the bet that Tenstorrent (Jim Keller's new crowd) are making.
Tenstorrent's cores are very special-purpose and infused with dedicated SRAM. That makes it a poor candidate for insight into general computing trends.
> it's hard for generic compute to compete against specialized HW > once the problem domain is well-enough specified...
No, they're targeting generic computing tasks with their E-cores, and providing a range of other solutions for special-purpose workloads. Everything from on-die AMX to Altera FPGAs, GPUs, and Habana's & Movidius' AI accelerators.
> for Intel to launch, say 8P- + 24E-cores is a hell of a bet
Agreed. It's hard to see the need for that much concurrency, in a mainstream desktop platform. You're right that we'd need a minor software revolution to put them all to good use. I have some ideas about how the programming model would have to shift, and it represents a significant departure from industry standard techniques, at both the user space and kernel level. That's obviously not going to happen in a mere couple of years.
> Transputers
I loved reading about them, back in the early 90's. Then, I had to program arrays of SHARC DSPs, their spiritual descendant, and it pretty much sucked. But, that had a lot to do with there being no global memory, not to mention cache coherence. So, the connection to modern E-cores is tenuous at best.
ccNUMA didn't really come onto the scene until after that.
"It remains unclear that E-core is a great match for the desktop." If by "desktop" you mean several (4-8) cores running hardly parallelizable workloads, then an E-core might be quite useful in laptops (with low energy use means both saving "direct use" electricity and "no cooling fans" electricity). Also, not using much of the power budget means that a performance core will have a lot more headroom for the couple of seconds it might be needed. Also, E-cores will be quite useful if you can launch lots and lots of threads (video processing, some engineering applications, ...).
But, if what you need is gaming (or similar performance profiles), then 4-8 big cores are all that you need.
> if what you need is gaming ... then 4-8 big cores are all that you need.
Yes, but not for the reasons you state. Games can scale graphics a lot more easily than they can scale what the CPU is mostly doing. Because of that, games tend to end up with a narrower window of CPU requirements than GPU requirements (i.e. in terms of relative performance). So, even if a game engine *could* find enough useful work to keep 32 E-cores busy, doing so would mean those gamers with a mere quad core CPU would be out of luck.
> E-cores will be quite useful if you can launch lots and lots of threads (video processing
Yes, though video processing is best done on GPUs.
I think a sea of E-cores would be great for building large software packages, however. That already scales well to CPUs with 32-cores and above.
they didn't really dial back e cores. they had few performance cores to start with and only upped core counts to match competition. their cores are already smaller than the competition as well thanks to the node advantage so there's no need to improve performance/area.
it's when you get to parity that the performance/area benefits come into play, but Apple pays more to get early access to next gen nodes so there's no way to compare designs directly until other designers are on the same node. they also control their whole software stack so you can't even compare performance figures as they vary wildly depending on compatibility with the M1.
> they also control their whole software stack > so you can't even compare performance figures > as they vary wildly depending on compatibility with the M1.
No, they don't control the apps. And there are some compute-intensive apps with native versions for both M1 and x86 (on both Apple and other platforms).
I think the reason Apple didn't scale E-cores is that for lower core-counts, it's more important to have P-cores. Just look at the Alder Lake desktop product stack. Once you have about 8 P-cores, you reach a point where any more cores is about increasing throughput rather than reducing latency. And that's where E-cores shine.
I'd imagine Mac Pro could do something like pairing 16 or 32 P-cores with 64 or 128 E-cores. I think it'd be cute if they'd put the Ultra on a PCIe card and use it as a dGPU. What does the Ultra have for PCIe connectivity?
I think you're being generous to imply Intel chose not to use more than 8 P cores. I'd say they are forced to keep it to 8 or less for power and thermal reasons on the consumer platform where they are forced to run on the wrong side of the frequency curve in order to compete. Likewise I see the hybrid design as being forced on them to compete on both single and multi thread fronts. Without that they wouldn't have been relevant in the mid to upper range this generation.
> Intel .. are forced to keep (P-cores) to 8 or less for power and thermal reasons
They absolutely *could* have used more P-cores, but the power & thermal limits would've meant decreasing the all-core clock speeds. And, at that point, the performance margins would've tilted even more in favor of E-cores.
Whether you look at perf/W or perf/area, E-cores are simply a better way to scale performance to highly-threaded workloads. I haven't seen any credible argument against that. The main concerns seem to be about impacts on thread scheduling - particularly as it pertains to more lightly-threaded workloads.
I would be curious to know if there is any update on when Intel plan using spintronics (especially MRAM, SOT-MRAM) as Persitent Memory in High Volume Manufacturing (HVM) ?
I wish so much to see new innovation spurring from MRAM cache and, ideally as DRAM replacement in SoC, maybe for Qualcomm 18A SoC…
It would bring the possibility to create new « always-on / normally off » innovative chips / products…
I’m not ready to believe in intels magic unicorns until I see a regular unicorn in real life. Let’s see chips fabbed on the process formerly known as 7nm, manufactured in quantity and sold in products people want to buy
Intel said tests chips for the Meteor Lake compute tiles were successful in Oct 2021. It's more likely TSM N3 GPU tiles could hold up that product, based on several reports of TSM N3 delays.
So Intel totally could fab GPUs on a process equivalent to TSMC's 3nm, but they can't, because they're fabbing so many other things on their process that is equivalent to TSMC's 3nm? What's the name of that process? And what are those other things being fabbed on it and where can I buy them?
What point are you trying to make exactly? Intel Arc uses TSMC as there wasn't enough fab capacity at Intel for both their CPU line and to fully enter the dGPU space.
MeteorLake, launching in 2023, is Intel's first disaggregated architecture (sorta like chiplets, but different.) - The bulk of the CPU will be built on Intel 4, with the Arc based, Battlemage iGPU tile will be manufactured on TSMC N3, since TSMC is the supplier for the Arc dGPU division.
As Intel is late with its process improvement, their only choice was to make old processors at the same size, or new processors with more transistors on the same process but at a higher physical size (so fewer processors per wafer, so fewer processors per fab per month). As Intel's process improvement over the past 10 or so years went without a hitch, Intel did not have spare fab capacity. The result - the same fabs could make fewer "2021" processors than "2019" processors. So, when the need to ramp up dGPU capacity appeared, Intel had no local capacity available and had to purchase from somewhere else.
> The result - the same fabs could make fewer "2021" processors than "2019" processors.
No, because their GPUs were never intended to be built on the 14 nm fabs where your statement applies. The real reason is that their 10 nm node(s) took too long to become competitive, resulting in insufficient time for them to build out their "Intel 7" fab capacity.
The issue of capacity is the same, but it stems from delays rather than inflated core counts, because I think they always planned to offer more cores on this node.
How does power delivery from the botom ened help ? Is it shorter leads, less capacitance ? Is it favorable when compared to the complexity increase such stack would require ?
It does lower power somewhat, yes, but the primary advantage is that it's become easier to shrink transistors than to shrink the wiring that connects them, so some designs are constrained by the density of wiring more than the density of transistors. Moving power delivery to below the logic frees up some of this congestion (as does the next step, moving the clock below the logic).
BUT (and this is always the case with Intel...) their obsessive crowing about this is very much more marketing than engineering. The designs that are primarily metal-constrained are the super-dense designs of Apple and other mobile chips, not so much Intel's (or for that matter AMD's) designs, which run at about the third of Apple's density. So when TSMC delivers backside power delivery, it will have a substantial effect on Apple's designs, but when Intel delivers backside power delivery little will actually change except a small power reduction.
> when Intel delivers backside power delivery > little will actually change except a small power reduction.
We should keep in mind that some of Intel's messaging is for the benefit of its foundry business and not necessarily tied to their x86 products. On that point, they're set to start producing some SoCs incorporating SiFive cores, later this year. They also appear to be offering them on the menu of IP for other IFS customers to use.
what a best type of article that I really appreciated, I just want to say thanks for sharing knowing all about the best info updates of the blog that looks very impressive. <a href="http://www.newcastletreeservicepros.com.au/"&...
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
89 Comments
Back to Article
Wereweeb - Monday, April 11, 2022 - link
It's a bit disappointing that Intel seems to have gotten it's shit together. I would have preferred to see them suffer for a little while longer, so that AMD can grow enough that we can be sure that they'll survive Intel's eventual rebound.But I guess it's still better than having Intel be burned so badly that the investors chose to sell off their foundry business, or stop making consumer CPU's, or something else along those lines. It's already a shame that everyone is talking about how they'll abandon Optane R&D.
In the end, we need someone else to make sure TSMC behaves, and Samsung seems to always be two steps behind them.
pipken21 - Monday, April 11, 2022 - link
Definitely welcome to see the competition light up. It can only be good for the end users to have more suppliers and options. I am a little skeptical that Intel will be able to pull off 4 processes on EUV in 2 years, but it will be impressive if they do.Samsung & TSMC have had years of experience with EUV but both still seem to have some issues with their latest processes.
JayNor - Monday, April 11, 2022 - link
Intel research on their EUV proceeded in parallel with the DUV 10nm multi-patterning.dotjaz - Sunday, April 17, 2022 - link
And you think TSMC&Samsung's EUV just sprung into life without any R&D done in parallel with their DUV? What do you think N7&8LPP are?Calin - Wednesday, April 13, 2022 - link
"Definitely welcome to see the competition light up. It can only be good for the end users to have more suppliers and options."Unfortunately, the presence of Intel in the market usually brings fewer suppliers and options - see Intel's past actions, some which generated fines and some which were settled out-of-court.
TheinsanegamerN - Sunday, April 17, 2022 - link
AMD had control of the consumer market for a scant 2 years before jacking up prices to the moon. Intel's near monopoly from 2007-2016 prices remained mostly stable at roughly $300-350 for high end consumer chips.Granted, they were quad cores, but still, a monopoly under either company is horrible. I dont want AMD getting any lazier then they already have.
Qasar - Sunday, April 17, 2022 - link
so 50 buck increase on MSRP is considered " jacking up prices to the moon. " yea ok sure.IF you are referring to the prices in the store, they werent all that bad where i am, picked up my 5900X for $699 cdn i think it was last year
intels prices went up a lot more then amds here before zen was released, and then the top end intel cpu dropped $1k shortly after that release.
between intel and amd, amd so far, as been the lesser evil of the 2.
Samus - Monday, April 18, 2022 - link
You have always been able to get reasonably capable AMD CPU's with Zen architecture and integrated graphics as low as $60 (Athlon 3000G) for years now, while powerful entry-level CPU's like the Ryzen 3200G had been selling around $120 prior to the pandemic (you'd have to go for something without integrated graphics to get in that price range now)This is all on par with Intel at the low-end. However, at the low-end, Intel is faster. AMD isn't competitive until you spend $150-$200. That really comes down to two things: for AMD to turn a profit with their smaller volume, they must sell more expensive chips, and the obvious: AMD is dependent on TSMC, who dictates supply, and to some extent, cost.
mode_13h - Monday, April 18, 2022 - link
Oh, @Samus... that's *far* too sane reasoning for an "insanegamer". :Dlizazampa - Friday, May 27, 2022 - link
Sonic’s menu includes the most traditional aspects of fast-food cuisine: hot dogs, chilli dogs,onion rings, soft drinks, slushies, milkshakes, hamburgers, french fries, corn dogs,
sundaes and banana splits.
They have continuously updated their fast food menu to satisfy their customers’ requirements.
mode_13h - Wednesday, July 13, 2022 - link
Well, that's a form of spam I haven't seen.thestryker - Monday, April 11, 2022 - link
The fab business being spun off is something some investors were already calling for so I think it's fair to say this is far more good than bad.AMD seems to be in a pretty solid position, even against SPR, and they've been making a lot of big moves so as long as none of it blows up I think they're good even with the imminent Intel rebound. It seems like each company is carving out their own priorities in the high profit sectors, and given production limitations (see TR Pro 5xxx supply issues) I think it'll be a few years before they're stomping on each other's toes again.
JayNor - Monday, April 11, 2022 - link
AMD has no announced CPU with HBM in package, as already sampled on SPR-HBM.mode_13h - Tuesday, April 12, 2022 - link
Yes, but 3D cache puts a big dent in HBM's advantage (for some workloads).Has Intel even announced the programming model supported by SPR-HBM? Other than using it as L4 cache, I think it's not going to see a lot of use for the first year or so, while software support begins to ramp up for it. I see SPR-HBM as a software development vehicle for hierarchical memory, more than anything else. Sure, there will be some niche applications that benefit quite nicely from it.
thestryker - Tuesday, April 12, 2022 - link
I don't believe they have announced the programming model, and the first chunk of them are going straight to HPC from what I've been able to find out. I'd imagine these will be the last SPR models to hit the market if they do at all.whatthe123 - Wednesday, April 13, 2022 - link
the hbm is probably on there for AVX512/AMX throughput in things like machine learning, not really displaced by milan-x. it'll end up competing with genoa.mode_13h - Wednesday, April 13, 2022 - link
> hbm is probably on there for AVX512/AMX throughput in things like machine learningAgreed. AMX / deep learning will need a lot of memory bandwidth. However, I think that's only part of the story.
> it'll end up competing with genoa.
Presumably, there'll be a Genoa-X, which will help against the case where Sapphire Rapids is simply using HBM as L4 cache.
name99 - Monday, April 11, 2022 - link
'Intel seems to have gotten it's shit together"Seems is the operative word. I'm sure Intel will have a node CALLED 18A. And they will ship something on it in 2023Q3. But will it actually be a significant change from the earlier node? And will the item shipped be of any relevance?
intel Arc is an interesting test case for the supposedly new, supposedly honest Intel.
That was going to ship Q1 2022. And, supposedly, in theory, it shipped, possibly, somewhere in the world, on March 30 which I guess is technically Q1 2022.
But can you actually buy one? And if you could, would you want to?
Now apply that same outcome to this process roadmap/fantasy...
eek2121 - Monday, April 11, 2022 - link
Whatever it may be CALLED, it will be competitive with a similar TSMC offering. The whole reason they renamed things in the first place was because they were the only ones adhering to such tight standards. 'Intel 7' is quite similar to TSMC N7, and Intel 4 is expected to be better than the TSMC '4nm' equivalent.Now that they are getting their process issues straightened out, they need to improve their Core design to address power efficiency concerns.200+W for a CPU is NOT cool.
michael2k - Tuesday, April 12, 2022 - link
They renamed it as a marketing exercise. If Intel's 10nm Enhanced SuperFin (it’s also been called 10nm+++ and 10nm++) was a bear to communicate.Also if it were competitive with TSMC N7 there wouldn't be a discussion about their competitiveness. Intel 7 has been in production since fall 2020; Apple was using TSMC N7 in fall 2018, two years earlier.
Now if you're saying Intel has finally caught up to N7P and that later this year they will catch up to N5, and finally be current with TSMC's N4 by next year... you're guessing as much as anyone else.
mode_13h - Tuesday, April 12, 2022 - link
> they will ship something on it in 2023Q3. But will it actually be a significant> change from the earlier node? And will the item shipped be of any relevance?
Yes, key questions. Many of us remember how they seemed to ship a few trays of Cannon Lake, just so they could tell investors they delivered on some revision of their 10 nm timeline.
> intel Arc ... That was going to ship Q1 2022.
Yes, March 30th was suspiciously close to the end of the quarter. Funny enough, "Intel shill" site Anandtech didn't even cover their launch announcement, unlike many other tech sites.
> Now apply that same outcome to this process roadmap/fantasy...
Comparing manufacturing process to a graphics product is not really apples-to-apples.
For me, the elephant in the room is still Intel's 10 nm debacle. That's reason enough to doubt everything Intel says about their manufacturing tech.
Calin - Tuesday, April 12, 2022 - link
"I'm sure Intel will have a node CALLED 18A"Intel was pretty honest until now with the 14nm, 14+, 14++, ...
wumpus - Saturday, May 7, 2022 - link
They may have been more honest about 14nm, 14+, 14++ but their schedules and roadmaps were pure fantasy for years (going back to the transition to 14nm). And this article is all about the "new improved marketing" of the schedule.Spunjji - Wednesday, April 13, 2022 - link
I remain sceptical for similar reasons.JayNor - Monday, April 11, 2022 - link
Intel responded to the Optane rumors. Gen 3 optane is still on its way. They also added optane support to the new agilex-M fpga. They also indicated plans to create a cxl controller that handles optane.mode_13h - Tuesday, April 12, 2022 - link
> Gen 3 optane is still on its way.Glad to hear it. I'll be even happier to see it! I still fancy the idea of an Optane boot drive, someday.
thestryker - Tuesday, April 12, 2022 - link
Gen 3 Optane will unfortunately be using the Gen 2 3D XPoint media, but with new control options. I too cross my fingers that there will be more generations forthcoming with more layers of cells as I know that's been a primary cost factor. There's a lot going for the technology, but unfortunately a lot of barriers for another client segment, but I'm still holding out hope.Mr Perfect - Monday, April 11, 2022 - link
Oh, I dunno, AMD was already getting to the point where they needed a good competitor again. Look at how long it took them to finish the Zen 3 product stack! It wasn't until Alder Lake that they took low to mid range parts seriously again. Not to mention how Zen 3 will probably get to celebrate it's second birthday before Zen 4 really kicks off.mode_13h - Tuesday, April 12, 2022 - link
> AMD was already getting to the point where they needed a good competitor again.I doubt it. I'm sure zero people at AMD were kicking back and snoozing before Alder Lake.
It's hard to appreciate how much work it is to scale up your customer base, but that's exactly what AMD has been doing, the past few years. Each one of those customers takes internal resources, wants special features, finds new bugs, etc. Rapid growth of a business is actually hard to do well, especially without compromising your execution.
> Zen 3 will probably get to celebrate it's second birthday before Zen 4 really kicks off.
Lots of things have to align, for a successful launch. Remember how Zen 3 used the same manufacturing process as the Zen 2 XT processors and the same platform. Zen 4 is on a new process node and new socket.
Also, try to remember that Zen+ was almost a direct port of Zen 1 to 12 nm. With that in mind, the gap between Zen 1 and 2 was also pretty big.
TheinsanegamerN - Sunday, April 17, 2022 - link
>i doubt itSo...did you just sleep through all of 2020 and 2021? Did you miss AMD jacking up their 6 core's effective price by 90% for a 20% uplift? The whole "no zen 3 below $270" thing that lasted until alder lake launched, then suddenly BAM price drop BAM price drop BAM new products?
Seems pretty obvious to me that AMD's old attitude of falling asleep at the wheel the moment they get the lead, which destroyed them with the first gen phenom, the HD 6000 series, and GCN 1.2, is alive and well.
Qasar - Sunday, April 17, 2022 - link
" So...did you just sleep through all of 2020 and 2021? " did you ? or did you not see what the world went though, and in still kind of going through? you DO realize we have and are still dealing with Covid-19, right ?mode_13h - Monday, April 18, 2022 - link
> then suddenly BAM price drop BAM price drop BAM new products?When Intel launches something that's actually competitive, the logical expectation would be for AMD's demand to slacken. And that lets them justify re-pricing and adding lower product tiers.
Can you imagine what would've happened if AMD did that a couple years ago? The products would've still sold out and most end consumers would've still paid similar prices, but the difference would've just gone to retailers and scalpers.
> AMD's old attitude of falling asleep at the wheel the moment they get the lead
Zen 4 will tell us if they took their foot off the gas. Until then, it'll be interesting to see how the 5800X3D benches in the real world. Even if it comes with heavy tradeoffs, I look at that as a learning experience, for them. I'm sure the 3D chiplets in following generations will benefit from the experience.
> first gen phenom, the HD 6000 series, and GCN 1.2, is alive and well.
In some cases, AMD stumbled into financial problems and had to make painful layoffs. In other cases, the competition just came back stronger. I'm not saying they didn't also have an under-performing mindset, which you can sort of see from the interviews with Jim Keller.
usiname - Tuesday, April 12, 2022 - link
How many years we got skylake refresh? AMD still make good improvement, not every year, but at least on 2, and Zen4 is expected to delivery 25-30% ipc imrpovement.TheinsanegamerN - Sunday, April 17, 2022 - link
Yeah they were making 20-25% gains per gen and the moment they passed intel the prices started going up, in the 6 core's case, 90% in a single gen for that 20% uplift.Never forget, the 3600 was a $169 product replaced with a $299 product. If you were a budget builder, go pound sand and buy intel.
Qasar - Sunday, April 17, 2022 - link
as my previous post mentioned, we are still going through a virus....mode_13h - Monday, April 18, 2022 - link
> the 3600 was a $169 product replaced with a $299 product.No, the 3600 launched at $199. It wasn't replaced until the 5600, which recently launched at... wait for it... $199!
The 3600X launched at $249 and the 5600X launched at $299.
As @Qasar, points out, that launched in Nov 2020, when demand for CPUs was abnormally high and Intel was still pushing Skylake cores on 14 nm. So, the $50 price increase was understandable, if unpalatable.
You're making at least two errors, here. You're equating 6-core with 6-core, regardless of positioning, and you're comparing the discounted *old* model against pricing of the new model. 3600 didn't *launch* at $169.
We could just blame supply problems & demand spikes on AMD and pat ourselves on the back for our moral outrage, but if you want more insight into *why* the prices shifted, then you can't ignore how the context changed.
Spunjji - Wednesday, April 13, 2022 - link
The Zen 3 product stack has been waiting for demand to slacken. AMD could have released these chips sooner, but they'd have only been cutting off profits for themselves.We don't have to /like/ it as buyers, but it's a cold calculus, and if we want AMD to survive as a competitor then some number of decisions like this will have to be made.
mode_13h - Wednesday, April 13, 2022 - link
> The Zen 3 product stack has been waiting for demand to slacken.I'm not sure about that. Zen 4 needs TSMC's N5 to be sufficiently mature (not to mention capacity) and DDR5 to have good availability. AMD should've wanted to launch it in Q1 of this year, especially if you consider all the discounting they're doing on the Ryzen 5000 product stack.
> AMD could have released these chips sooner
So, I reject the assertion that AMD simply didn't *want* to launch Zen 4 any sooner than they are. I think they're getting it out the door as soon as they reasonably could.
TheinsanegamerN - Sunday, April 17, 2022 - link
>The Zen 3 product stack has been waiting for demand to slacken. AMD could have released these chips sooner, but they'd have only been cutting off profits for themselves.Isnt it funny that it just so happens the demand slackens as SOON as intel's alderlake goes on sale? Almost like that's a massive cope.
mode_13h - Monday, April 18, 2022 - link
> Isnt it funny that it just so happens the demand slackens as SOON as intel's alderlakeNo, it's perfectly logical. Alder lake is much more competitive against Zen 3 than either generation before it. Of course that's going to steal some of Zen 3's demand! We'd be fools to pretend otherwise.
mode_13h - Tuesday, April 12, 2022 - link
You're so concerned with competition among CPU makers, but what about ASML? In particular, their exclusive partnership with Intel on high-NA (as in aperture -- not sodium!) is a little troubling.> it's still better than having Intel be burned so badly that
> the investors chose to sell off their foundry business
I still think that would be for the best. I'm glad Intel is starting to treat it more as a separate business, since that will ideally be the eventual outcome.
Calin - Tuesday, April 12, 2022 - link
"18A development has been moving so well that the company’s R&D operations are now on or ahead of all of their development milestones"That does not mean that they will reach risk production as expected, or "real" production numbers as expected. It means that some hard problems were solved, and some other hard problems remain.
As for the additional space needed... improved processes usually take longer (or much longer) per wafer, and (due to the increase in transistor counts) the number of processors per wafer didn't increase much.
Papaspud - Tuesday, April 12, 2022 - link
AMD has shown that when they get on top- they are just as bad as Intel. Look at their latest offerings= r5 5500 4500 and radeon 6500, all total trash products.mode_13h - Wednesday, April 13, 2022 - link
> AMD has shown that when they get on top-So, you think the determining factor was AMD's recent dominance? I say it's the historic chip shortage.
The RX 6500 was designed to be cheap and crypto-unfriendly by cutting die size and memory. The Ryzen 5500 and 4500 are probably just selling off some low-binned APUs at a discount. Right now, I think it's difficult for AMD to provide a better answer to Intel, at the entry level.
Spunjji - Wednesday, April 13, 2022 - link
Some people want AMD to keep offering more than Intel for less money, just so they can buy Intel and talk about how trash AMD's products are.Spunjji - Wednesday, April 13, 2022 - link
"all total trash products"Not really, though. You picked their weakest and cheapest products, all of which serve specific niches and all of which undercut their competition (dramatically in the case of the 6500 XT).
Tom Sunday - Friday, April 22, 2022 - link
Yes indeed in the end we all need someone else to make sure TSMC and likewise offshore friends are kept in check! With Taiwan now being seriously in the crosshairs by China and global superpower territorial expansionism is here to stay, I wonder what that will bring next to our own tech-doorstep? Taiwan as we speak is also not welcomed being a member of the U.N or even considered by many to be a sovereign nation. The good news however is that the USA is now being seen by many as a ‘Safe Harbor’ country. Intel here in Oregon across four campuses in Hillsboro – just ten miles west of Portland just opened its $3-billion expansion with close to 21,000 employees. To date over 65,000 job applications have been received and with people vying for the benefits such as 401K, unrestricted medical and dental, paid vacations, onsite chld care, guaranteed long-term fulltime employment and the top salaries Intel offers. Most certainly and ignoring the technical aspects discussed here for a minute all of this makes me a fan of Intel!wumpus - Saturday, May 7, 2022 - link
Intel will abandon Optane once they can no longer pretend it is dead.If it wasn't fully dead, Micron wouldn't have simply thrown their share of it away. This is rather sad as had it worked, it wouldn't have been a stretch to make huge DRAM caches for vast Optane "main storage". But it don't expect it to ever be more than a really expensive alternative to flash marketed only to enterprise customers.
Silver5urfer - Monday, April 11, 2022 - link
Well it was expected that R&D path that Gelsinger took Intel which it needs is showing. Good to see more R&D on the Lithography side. Especially when we have the total domination by TSMC. A balance needs to be there. Note Samsung is also in the race of GAAFET 3nm same as Intel 3 or 20A / TSMC 3 or 2.But unfortunately, all this tech is going to be wasted on the garbage BGA trash more than HEDT or Mainstream Desktop because that's where these investor riddled companies are at. And worse is E-Core BS that Intel created because of their Intel Core uArch being too old, Nehalem. So until 2025 if the Z990 or Z1000 series chipsets, the Intel Core branding will relegate itself to the pathetic ARM copy cat design of growing more E Cores trash. Maybe by 18A+ with NAEUV Intel will redesign their x86 with a successor to Core. Also hopefully no more BS of awful security holes and tons of papers from Technion students, I have a benefit of doubt how Intel Haifa behaves given the history with NSO Group and Pegasus feud with Apple corporation.
Plus their Xeon will give competition to Bergamo after a whole year, by 2023 Bergamo will be out, and 2024 is when Xeon with E Cores only will come. So Intel is decided on this big small path unfortunately so is AMD but AMD so far doesn't incorporate both into one CPU. I hope they do not do that for Zen 4 or any CPU on AM5. My guess is AMD will make 2 SKU classes, absolute top will be high performance Big trad x86 and the 2nd class would be c cut down versions.
Finally what's the use when the PC itself is seeing weird changes such as Windows 11 going more for BGA Touchscreen dumpster rather than poweruser based UX like Windows 7. And by 2025 and up, not sure how GaaS and these pathetic Service models will fare. So all this tech and growth wasted on such is a massive disappointment.
JayNor - Monday, April 11, 2022 - link
Intel will soon have 16 e-cores on the desktop chips ... Raptor Lake... and the rumors are that will increase to 24 e-cores on the chip after Meteor Lake.I see wccftech reporting a 56 core alder lake-x chip. Will that be mostly e-cores?
usiname - Tuesday, April 12, 2022 - link
With tdp of 300w I dont see how this cpu will have more than 8p cores and 48e coresGondalf - Wednesday, April 13, 2022 - link
If they clock lower with an high single core turbo your 300 W will be an idiocy. More cores, lower average clock speed.FunBunny2 - Monday, April 11, 2022 - link
"Finally what's the use when the PC itself is seeing weird changes such as Windows 11 going more for BGA Touchscreen dumpster rather than poweruser based UX like Windows 7."there was a time, some years ago, when some wag (no longer remember who; might have been Dvorak) allowed as how there were three killer apps for the PC
- spreadsheets
- word processing
- spreadsheets
today we might add a fourth
- email
either way, a 486 with a decent bit of wifi will do for most folks.
mode_13h - Tuesday, April 12, 2022 - link
> today we might add a fourth> - email
When was that written? 1995?
Web, online shopping, and social media are obviously killer apps. Sure, you can do them on a phone, but they're all better on a PC.
And that's leaving out the granddaddy of them all, which is gaming. There are gaming and education apps that don't port well to consoles or phones.
> a 486 with a decent bit of wifi will do for most folks.
No. The web is a moving target. What CPU power it took to brows the web of the 1990's over dialup (or a 1 mbps cable modem, for the lucky) will no longer cut it. A Raspberry Pi is lightyears faster than any 486, and it's still painful for certain things.
I think this little troll post backfired on you. It seems as if you remember nothing of the past 2 decades.
DanNeely - Tuesday, April 12, 2022 - link
That sounds vaguely familiar, and Dvorak was my only regular source of tech industry snark in the 90s so him being the author seems plausible.Calin - Tuesday, April 12, 2022 - link
486 is theoretically limited to 4 GiB of addressable memory. However, you would be hard pressed to find a mainboard supporting 64MiB, with 32 MiB a more credible alternative.So, while the 486 could be fast enough (which I doubt), it will most certainly not have enough memory.
mode_13h - Tuesday, April 12, 2022 - link
I had a Pentium Pro with 128 MB of RAM, but I forget whether it was FPM or EDO. I ran a few of my own benchmarks on it, and I think it managed about 90 MB/sec for read/write. On any 486, it should've been some fraction of that. So, even if you could pack it with say 2 GB of RAM, actually using it would've been another thing entirely.I think the first machine I had with 2 GB was a Pentium 4, which ran at 3.2 GHz compared to the 33 MHz of a typical 486.
lmcd - Tuesday, April 12, 2022 - link
Do you reflect at all before you post garbage like this? I half expected to see a Bengazi reference thrown in for good measure.kwohlt - Tuesday, April 12, 2022 - link
Once AMD goes big.Little in 2 years or so, and every CPU manufacturer, from Intel, AMD, Qualcomm, Mediatek, and Apple are all using big.little, will people who have no formal training or real world experience in CPU design still be calling the entire industry of competing companies and engineers 'idiots'?And you're mad Windows 11 accommodates touch? I like using a touch screen convertible device.
name99 - Tuesday, April 12, 2022 - link
He may be unbalanced, but he's not that wrong.It remains unclear that E-core is a great match for the desktop.
E-core clearly has massive value going downward (iPhone, iPad, laptop; but also Apple Watch and HomePod mini).
And E-core probably has a useful role in the data center for servers whose primary job is to move data around with minimal manipulation thereof (things like memcached servers).
But on the desktop? Look at how Apple dialed BACK the number of E-cores for M1 Pro and Max. It's surely significant that the company with the most experience of these things made that decision?
Now there is a possible direction E-cores could go, which is to provide a sea of generic throughput computing. This is essentially the bet that Tenstorrent (Jim Keller's new crowd) are making. And you could imagine such a sea of generic throughput being useful for NPU, GPU and ISP type applications. But that's a hell of a bet and it's unclear that it will pan out; Intel made a similar bet with the first version of Larrabee as GPU, and, well, we all know how that went -- it's hard for generic compute to compete against specialized HW once the problem domain is well-enough specified...
So for Intel to launch, say 8P- + 24E-cores is a hell of a bet that enough important SW companies figure out, in just a year or two, a way to extract value from a flock of chickens. The industry, more or less, has been working on this problem since at least Transputers and the BBN Butterfly, with limited success, so it seems unlikely that we'll see the relevant breakthrough soon.
mode_13h - Wednesday, April 13, 2022 - link
> Look at how Apple dialed BACK the number of E-cores for M1 Pro and Max.Apples and Oranges, I say. Apple's P-cores are a lot more efficient than Intel's, so Apple probably doesn't have to worry about thermal throttling with the number they were willing to pack onto the Pro and Max dies. And it's not like they have so many P-cores, either. Where E-cores really come into their own is when you start to scale to even higher core counts.
> And E-core probably has a useful role in the data center for servers whose primary
> job is to move data around with minimal manipulation thereof
Intel's E-cores aren't *that* weak. Gracemont is about 65% as fast at integer workloads as Golden Cove (both, single-thread), in Alder Lake. Once you scale up core counts, the P-cores would have to run at lower clocks, thereby narrowing the gap further.
Also, Sapphire Rapids will introduce special-purpose "DSA" engines for simple data movement tasks.
> a possible direction E-cores could go, ... a sea of generic throughput computing.
> This is essentially the bet that Tenstorrent (Jim Keller's new crowd) are making.
Tenstorrent's cores are very special-purpose and infused with dedicated SRAM. That makes it a poor candidate for insight into general computing trends.
> it's hard for generic compute to compete against specialized HW
> once the problem domain is well-enough specified...
No, they're targeting generic computing tasks with their E-cores, and providing a range of other solutions for special-purpose workloads. Everything from on-die AMX to Altera FPGAs, GPUs, and Habana's & Movidius' AI accelerators.
> for Intel to launch, say 8P- + 24E-cores is a hell of a bet
Agreed. It's hard to see the need for that much concurrency, in a mainstream desktop platform. You're right that we'd need a minor software revolution to put them all to good use. I have some ideas about how the programming model would have to shift, and it represents a significant departure from industry standard techniques, at both the user space and kernel level. That's obviously not going to happen in a mere couple of years.
> Transputers
I loved reading about them, back in the early 90's. Then, I had to program arrays of SHARC DSPs, their spiritual descendant, and it pretty much sucked. But, that had a lot to do with there being no global memory, not to mention cache coherence. So, the connection to modern E-cores is tenuous at best.
ccNUMA didn't really come onto the scene until after that.
mode_13h - Wednesday, April 13, 2022 - link
> Intel's E-cores aren't *that* weak. Gracemont is about 65% as fast at integer workloads> as Golden Cove (both, single-thread), in Alder Lake.
In fact, if we compute IPC by naively dividing each by their respective single-core turbo speeds, Gracemont rises to 86% of Golden Cove's IPC.
Source: https://www.anandtech.com/show/17047/the-intel-12t...
Calin - Wednesday, April 13, 2022 - link
"It remains unclear that E-core is a great match for the desktop."If by "desktop" you mean several (4-8) cores running hardly parallelizable workloads, then an E-core might be quite useful in laptops (with low energy use means both saving "direct use" electricity and "no cooling fans" electricity). Also, not using much of the power budget means that a performance core will have a lot more headroom for the couple of seconds it might be needed.
Also, E-cores will be quite useful if you can launch lots and lots of threads (video processing, some engineering applications, ...).
But, if what you need is gaming (or similar performance profiles), then 4-8 big cores are all that you need.
mode_13h - Wednesday, April 13, 2022 - link
> if what you need is gaming ... then 4-8 big cores are all that you need.Yes, but not for the reasons you state. Games can scale graphics a lot more easily than they can scale what the CPU is mostly doing. Because of that, games tend to end up with a narrower window of CPU requirements than GPU requirements (i.e. in terms of relative performance). So, even if a game engine *could* find enough useful work to keep 32 E-cores busy, doing so would mean those gamers with a mere quad core CPU would be out of luck.
> E-cores will be quite useful if you can launch lots and lots of threads (video processing
Yes, though video processing is best done on GPUs.
I think a sea of E-cores would be great for building large software packages, however. That already scales well to CPUs with 32-cores and above.
whatthe123 - Wednesday, April 13, 2022 - link
they didn't really dial back e cores. they had few performance cores to start with and only upped core counts to match competition. their cores are already smaller than the competition as well thanks to the node advantage so there's no need to improve performance/area.it's when you get to parity that the performance/area benefits come into play, but Apple pays more to get early access to next gen nodes so there's no way to compare designs directly until other designers are on the same node. they also control their whole software stack so you can't even compare performance figures as they vary wildly depending on compatibility with the M1.
mode_13h - Wednesday, April 13, 2022 - link
> they also control their whole software stack> so you can't even compare performance figures
> as they vary wildly depending on compatibility with the M1.
No, they don't control the apps. And there are some compute-intensive apps with native versions for both M1 and x86 (on both Apple and other platforms).
I think the reason Apple didn't scale E-cores is that for lower core-counts, it's more important to have P-cores. Just look at the Alder Lake desktop product stack. Once you have about 8 P-cores, you reach a point where any more cores is about increasing throughput rather than reducing latency. And that's where E-cores shine.
I'd imagine Mac Pro could do something like pairing 16 or 32 P-cores with 64 or 128 E-cores. I think it'd be cute if they'd put the Ultra on a PCIe card and use it as a dGPU. What does the Ultra have for PCIe connectivity?
Jp7188 - Wednesday, April 13, 2022 - link
I think you're being generous to imply Intel chose not to use more than 8 P cores. I'd say they are forced to keep it to 8 or less for power and thermal reasons on the consumer platform where they are forced to run on the wrong side of the frequency curve in order to compete. Likewise I see the hybrid design as being forced on them to compete on both single and multi thread fronts. Without that they wouldn't have been relevant in the mid to upper range this generation.mode_13h - Thursday, April 14, 2022 - link
> Intel .. are forced to keep (P-cores) to 8 or less for power and thermal reasonsThey absolutely *could* have used more P-cores, but the power & thermal limits would've meant decreasing the all-core clock speeds. And, at that point, the performance margins would've tilted even more in favor of E-cores.
Whether you look at perf/W or perf/area, E-cores are simply a better way to scale performance to highly-threaded workloads. I haven't seen any credible argument against that. The main concerns seem to be about impacts on thread scheduling - particularly as it pertains to more lightly-threaded workloads.
drothgery - Tuesday, April 12, 2022 - link
And by "garbage BGA trash" you mean the U series CPUs that power a large majority of shipping PCs and have for quite a long time now?Diogene7 - Monday, April 11, 2022 - link
I would be curious to know if there is any update on when Intel plan using spintronics (especially MRAM, SOT-MRAM) as Persitent Memory in High Volume Manufacturing (HVM) ?I wish so much to see new innovation spurring from MRAM cache and, ideally as DRAM replacement in SoC, maybe for Qualcomm 18A SoC…
It would bring the possibility to create new « always-on / normally off » innovative chips / products…
Slash3 - Monday, April 11, 2022 - link
MRAM will always be two years out, from whatever the current date is.Blastdoor - Monday, April 11, 2022 - link
I’m not ready to believe in intels magic unicorns until I see a regular unicorn in real life. Let’s see chips fabbed on the process formerly known as 7nm, manufactured in quantity and sold in products people want to buyJayNor - Monday, April 11, 2022 - link
Intel said tests chips for the Meteor Lake compute tiles were successful in Oct 2021. It's more likely TSM N3 GPU tiles could hold up that product, based on several reports of TSM N3 delays.Blastdoor - Tuesday, April 12, 2022 - link
Blaming TSMC for Intel not getting a product out on time -- that's something I never expected to read.Gee, I wonder why Intel isn't using their own fabs for the GPU...
onewingedangel - Tuesday, April 12, 2022 - link
Capacity - as long as everyone is capacity constrained and their own fabs are fully booked they want to drink AMD's milkshake as well.Blastdoor - Tuesday, April 12, 2022 - link
So Intel totally could fab GPUs on a process equivalent to TSMC's 3nm, but they can't, because they're fabbing so many other things on their process that is equivalent to TSMC's 3nm? What's the name of that process? And what are those other things being fabbed on it and where can I buy them?kwohlt - Tuesday, April 12, 2022 - link
What point are you trying to make exactly? Intel Arc uses TSMC as there wasn't enough fab capacity at Intel for both their CPU line and to fully enter the dGPU space.MeteorLake, launching in 2023, is Intel's first disaggregated architecture (sorta like chiplets, but different.) - The bulk of the CPU will be built on Intel 4, with the Arc based, Battlemage iGPU tile will be manufactured on TSMC N3, since TSMC is the supplier for the Arc dGPU division.
Calin - Wednesday, April 13, 2022 - link
As Intel is late with its process improvement, their only choice was to make old processors at the same size, or new processors with more transistors on the same process but at a higher physical size (so fewer processors per wafer, so fewer processors per fab per month).As Intel's process improvement over the past 10 or so years went without a hitch, Intel did not have spare fab capacity. The result - the same fabs could make fewer "2021" processors than "2019" processors.
So, when the need to ramp up dGPU capacity appeared, Intel had no local capacity available and had to purchase from somewhere else.
mode_13h - Wednesday, April 13, 2022 - link
> Intel did not have spare fab capacity.Yes.
> The result - the same fabs could make fewer "2021" processors than "2019" processors.
No, because their GPUs were never intended to be built on the 14 nm fabs where your statement applies. The real reason is that their 10 nm node(s) took too long to become competitive, resulting in insufficient time for them to build out their "Intel 7" fab capacity.
The issue of capacity is the same, but it stems from delays rather than inflated core counts, because I think they always planned to offer more cores on this node.
Jp7188 - Wednesday, April 13, 2022 - link
Eh? Intel chose TSMC because they couldn't make a competitive product with an in-house process. Intel themselves was pretty open about that fact.GC2:CS - Tuesday, April 12, 2022 - link
How does power delivery from the botom ened help ? Is it shorter leads, less capacitance ?Is it favorable when compared to the complexity increase such stack would require ?
name99 - Tuesday, April 12, 2022 - link
It does lower power somewhat, yes, but the primary advantage is that it's become easier to shrink transistors than to shrink the wiring that connects them, so some designs are constrained by the density of wiring more than the density of transistors. Moving power delivery to below the logic frees up some of this congestion (as does the next step, moving the clock below the logic).BUT (and this is always the case with Intel...) their obsessive crowing about this is very much more marketing than engineering.
The designs that are primarily metal-constrained are the super-dense designs of Apple and other mobile chips, not so much Intel's (or for that matter AMD's) designs, which run at about the third of Apple's density. So when TSMC delivers backside power delivery, it will have a substantial effect on Apple's designs, but when Intel delivers backside power delivery little will actually change except a small power reduction.
mode_13h - Wednesday, April 13, 2022 - link
Thanks for the explanation.> when Intel delivers backside power delivery
> little will actually change except a small power reduction.
We should keep in mind that some of Intel's messaging is for the benefit of its foundry business and not necessarily tied to their x86 products. On that point, they're set to start producing some SoCs incorporating SiFive cores, later this year. They also appear to be offering them on the menu of IP for other IFS customers to use.
* https://www.anandtech.com/show/16777/intel-license...
* https://www.anandtech.com/show/16780/intel-to-crea...
Lord of the Bored - Wednesday, April 13, 2022 - link
So we're all here to look at Intel's big D1X, then?cowymtber - Sunday, April 17, 2022 - link
Four nodes, in three years....from Intel. If you believe this will happen, then you probably believe in the Easter Bunny.mode_13h - Monday, April 18, 2022 - link
I don't know if the Easter Bunny is real, but I sure do like these little brown "eggs" it leaves behind!; )
genbel - Thursday, April 21, 2022 - link
what a best type of article that I really appreciated, I just want to say thanks for sharing knowing all about the best info updates of the blog that looks very impressive. <a href="http://www.newcastletreeservicepros.com.au/"&...mode_13h - Friday, April 22, 2022 - link
spammer.