Thanks for your coverage of fanless mini-PCs. However, I really wish you'd include something with "big cores", so we can get a sense of the scale of performance difference between them and Tremont.
Another nice-to-have would be at least a few benchmarks including a Raspberry Pi 4. However, it has serious thermal throttling issues, unless it's actively cooled or you use a substantial passive cooling solution.
I guess the ideal comparison would be a Tiger Lake-based system, since that's the same vintage and similar manufacturing tech as Tremont. Probably much harder to find in a fanless mini-PC, unless we're talking about an industrial PC, but I'd love even to see a comparison between two NUCs: Tiger Lake vs. Tremont.
Yes, Tiger Lake NUCs were made, but also very hard to come by: I have both.
In a way they are perfect to showcase the benefit of E/P cores …in the case of Intel: AMD is really another story.
The two NUCs look nearly identical on the outside, but inside they are very different beasts.
For starters: The Tiger Lake NUC11 (i7-1165G7 with 96EU Xe iGPU) is configured with a 64 Watt PL2, a rather long TAU and even the PL2 is 30 Watts by default, I believe. There is a reason it comes with a 90 Watts power brick! I changed PL2 to 50, TAU to 10 seconds and PL1 to 15 Watts to ensure the fan would never howl they way it does with the defaults.
I’ve seen HWinfo report a 5GHz maximum clock, but 4.7GHz is the official top speed. It’s at 64 Watts and near 5GHz clocks that I have measured 1707/5808 Geekbench 4 results on Linux (always a bit faster than on Windows). Jasper Lake doesn’t quite play in the same league at 781/2540 using 3.3 GHz and 10 Watts. In Watts/compute power Tiger Lake looks rather worse than Jasper Lake, but when it comes to rendering a complex web page or recalculating a giant Excel sheet, its sprinting power certainly has it appear much faster.
At 64 Watts the Tiger Lake is a desktop CPU, shoehorned into mobile power envelopes. And when it’s constrained to the levels that passive cooling can manage (see the Supermicro SYS-E100-12T-H review here), it really struggles to deliver that performance. The great thing about the Tiger Lake NUC is that you can change PL1, PL2 and TAU to pretty much anything you want and when you set it to the 10 Watts the Jasper Lake gets to use as an absolute maximum, it starts to do rather badly.
Some of that is because the iGPU always gets preference, leaving close to nothing to the CPU. But some of that is that the remaining power budget forces very low frequencies, where the big Core CPU loses against the Atom cores running at a full speed with these Watts.
Jasper Lake, like all the other Atoms since the J1900, never slows down. I’ve never seen it drop below its “Turbo” clock unless idle, even on a mix of Prime95 and Furmark, and I’ve never seen it exceed 10 Watts of combined CPU+GPU power consumption either.
I also have two Ryzen 5800U based notebooks (1443/7855 on Geekbench4), one of which can be switched between 15 and 28 Watts of TDP. When Tiger Lake and Zen 3 are strictly set to the same power levels, Tiger Lake has to run much slower even with half the cores: Ryzen beats it with a much smaller energy footprint per core. But with Tiger Lake left at the default NUC settings (which a battery powered notebook could not support), its four cores will beat an eight core Zen 3 at 15 Watts in Geekbench, which luckily never seems to exceed TAU.
Intel needs E/P because P cores need too much power at the clock rates they require to beat a Ryzen core, and only with E cores they can hit the efficiency of Zen cores in fully multi-threaded loads.
Wow, another awesome post! Thanks for taking the time to relate your findings. Very interesting!
> the iGPU always gets preference, leaving close to nothing to the CPU.
Very key point, but also one that Intel could conceivably address, to some extent, in future BIOS updates. Not that they're likely to, if it had been on the market for a while when you tested, but it's conceivable.
> in Geekbench, which luckily never seems to exceed TAU.
Another great point! I have never run Geekbench myself, and I haven't noticed reviewers mention this key detail.
Benches I've seen suggest both are very similar in multi to a i3 low power Skylake eg a ie-6100T (2core 4 thread very common thin client chip) - the gfx capability also seems also a close match for the 24EU part [probably a very similar part with improved HEVC support] (the 32EU N6000 should be better)
For single threaded the old Skylake is ~+50% faster., and from Skylake to Alder Lake it's nearly 2x , so nearly 3x from N5100 to i5-12500 for single thread
I have an old fanless Atom Z3735F (22nm) and these new SoCs are a impressive step up (~7x both cpu and gpu) -- I think the Pi Model B latest is very roughly 2x better than that nut no where near the 5100T in any metric.
tldr both benches would have been a wash one way of the other.
> tldr both benches would have been a wash one way of the other.
Huh? If old Skylake is 50% faster, and Jasper Lake is 3.5x as fast as Pi 4 Model B (which seems rather generous), then it wouldn't be "a wash", which is defined as:
I have a PI4 with 8GB of RAM in a metal case that supports a 2GHz overclock without active cooling: pretty much the best PI you can have these days.
I also have an Nvidia Tegra based Jetson Nano with 4GB of RAM.
At 2GHz the PI reaches 272/648 on Geekbench 4, the Tegra has to make do with 206/718 at 1.4GHz. The N6005 Jasper Lake reaches 781/2540 very similar to a Sandy Bridge i7-2600 at 3.8GHz Turbo.
The Jetson Nano actually does reasonably well on my 43" 4k desktop for basic 2D work, because it has a GPU with 128 Maxwell cores. Of course its CPU power is at the level of a Snapdragon 800 mobile phone.
The PI struggles badly at 4k, because the GPU has much less muscle. The slightly faster CPU is hard to notice.
Actually it was when Tom's hardware did a report on a PI compute cluster, that I wanted to retort just how stupid that project was, because you could get a single Jasper Lake Atoms for much less money, that would run rings around that cluster and could in fact simulate it all in software via VMs.
And that's when I found that finally a Jasper Lake NUC was available for purchase at €200 (including VAT) and immediately ordered one of the first and last ever sold here.
And yes, it runs rings around both with roughly 4x the CPU power, 64GB of RAM expandability and quite a reasonable GPU performance on a 4k display.
My favorite usability test is to use the "3D Globe View" on Google Maps under a Chrome based browser on Windows and to then tilt and turn a city landscape there. It's about the most efficient 3D graphics pipeline I've ever seen (puts Flight Simulator to total shame!) and performs quite reasonable on such a Jasper Lake NUC. With Firefox it's much worse on these low power devices, but with a beefy PC you'd never notice.
After quite a bit of tweaking I managed to get it to work on both the PI and the Tegra at 1920x1080 and the Tegra even gave a bit of interactivity thanks to its much stronger GPU. But on the PI that was about one frame a minute.
The PI and Nano are toys and ok for the €100 I spent on each.
A Jasper Lake NUC is quite a reasonable desktop machine and even an interesting micro server for some real workloads.
At €200 (without RAM or storage) the price/performance ratio is very hard to beat, but evidently none of the vendors really want you to know or buy that. I think it's the major reason you never could.
> At 2GHz the PI reaches 272/648 on Geekbench 4, the Tegra ... 206/718 at 1.4GHz.
Keep in mind that Jetson Nano has ostensibly 2x the memory bandwidth of the Pi v4. That surely helps offset the difference in raw CPU performance, as well as with 4k display performance.
Oh, and if that test was with the machines driving a 4k display, then merely refreshing your monitor will have been using a non-insignificant amount of the Pi's memory bandwidth (about 1 GB/s).
> N6005 Jasper Lake reaches 781/2540
Wow! Dual-channel memory configuration, I presume?
> on the PI that was about one frame a minute.
Uh... that sure sounds like you were using a software rendering path. The Pi's GPU is trash, but that's simply atrocious!
> evidently none of the vendors really want you to know or buy that. > I think it's the major reason you never could.
I'm reasonably confident it's actually just supply chain-related. Intel has been steering its limited fab capacity towards more profitable models and probably steering its limited supply of Jasper Lakes to chromebooks, where they're probably desperate not to lose market share.
there's like 5 different "brands" selling same thing on AliExpress etc. it runs win 10 just fine and is enough for 1080p Minecraft and basic office computing. great deal. most models have Intel 2.5G Ethernet too.
You could get SATA, but not PCIe. According to this, Jasper Lake and Elkhart Lake have only x8 PCIe 3.0 and x2 SATA ports.
Most boards are probably going to give you a x4 NVMe slot. Then, they could use a 3rd Party SATA controller to give you 4 more ports. Then, if they compromise on the bandwidth to that SATA controller, you can have a second Ethernet port and then a x1 PCIe slot that just might be open-ended (but probably not), to support a graphics card.
Sorry, but they really kneecapped this platform relative to what it could've been. You might do better with some equivalent Atom-branded CPUs. Atom C-series (Parker Ridge) has 16 integrated SATA ports, x32 PCIe 3.0 lanes, and up to 8 cores. P-series (Snow Ridge) has the same, but up to 24 cores.
The Surface Pro X from 2019 has a Microsoft SQ1 SoC, which is basically a Snapdragon 8cx and consists of 4x Kryo 495 Gold @ 3 GHz+ 4x Kryo 495 Silver @ 1.80 GHz (manufactured on TSMC 7 nm). According to wikichip, these are tweaked A76 and A55 cores. So, that seems credible, if not exactly an outcome I'd have presumed.
Something to keep in mind is that Jasper Lake is meant cheap chromebooks. Like, sub-$200 cheap, whereas Snapdragon 8cx is a premium part.
Snapdragon 7c (Gen 1?) should be more comparable in price to Jasper Lake. I think I've seen that as low as $170-200. Also, the Apcsilmic Dot 1 and ECS LIVA Mini Box QC710 mini PCs recently launched with the 7c starting at around $219.
If the leaks about Alder Lake-N are true, it will shake things up, if the price is right.
It will be a shrink from 10nm to Intel 7 (formerly 10nm Enhanced SuperFin), and they should have different cell libraries available. I think they'd want the densest one. I'm sure whatever it becomes will be impressive compared to Jasper Lake.
More importantly, I hope we see a lot of the 8-cores.
Of course 8cx devices are more premium. But I was not referring to price, but to what it is technically possible within a 7W power envelope if you include devices, which does not contain Intel CPUs. It just shows that Intel CPUs are incredibly power inefficient independent of price. This includes Lakefield, Jasperlake or whatever CPU Intel designed for the sub 9W TDP market.
Above, we were talking about how Jasper Lake compares with Skylake and ARM A76 cores. It would be fantastic to have some hard data on this. Why not run SPEC bench on these mini PCs? It can't be any less weird than running it on a iPhone, right?
A few typos here and there but a good review of hardware that might, for once, realistically land on someone's desk these days. Pity about the Liva's cooling situation. I wonder if chopping open a fair amount of the case top and replacing it with a fan grille or an epoxy-attached bit of non-metallic screen would be sufficient to mitigate the cooling problems while still retaining the bulk of the case itself. Too bad ECS didn't do potential buyers any justice with this design since the cost is otherwise okay.
I can’t help but wonder if Anandtech feels a certain affinity to near-death technology these days…
Jasper Lake is unfortunately dead. Whatever is being sold, was manufactured long ago and these systems are the typical leftover stock that Intel is pushing out the back door, as it reduces internal stockpiles to cover long-term warranties: nearly rock bottom, in other words.
I’ve been using Atoms since the J1900 on ASRock Mini-ITX boards, that were 100% passive, simply because they were 100% passive and thus zero noise: a somewhat novel experience ever since I switched from an Apple ][ [clone] to an IBM-AT [clone], which unfortunately started the trend on noisily moving parts.
And one constant has been, that Atoms have always supported way more RAM than advertised. It was 16GB or DDR3 on the DDR3 Atoms, 32GB for the J5005 devices I’ve been running as a oVirt(RHEV) cluster for a couple of years now.
A couple of weeks ago a Jasper Lake or Atlas Canyon NUC11 (NUC11ATKPE) popped up on my radar and since I’d been on a constant lookout for them, I immediately grabbed one, only to see that it might have been both, the first and last gasp of a phantom device: it’s been unavailable since, just as it was for it’s entire official life-time before.
And after putting it through its paces that is quite sad, because it really is a rather strappy and solid little machine, especially at a 2x Raspberry PI4 price point. It performs near identical to a Sandy Bridge i7-2600 and beats the Pi into a pulp with only a tiny helping of extra Watts: 10 vs 28nm does make a difference!
And again, just as you noticed, Jasper Lake will run just fine with far more RAM than Intel wants you to know. That is a constant with every NUC I’ve owned, i7 based NUC8/10/11 all run with 64GB just fine, even if only the Tiger Lake, from which I borrowed the dual 32GB DDR4-3200 SO-DIMMs for the Jasper Lake evaluation, did so officially.
But where my Tiger Lake will squeeze 40GB/s from that RAM, Jasper Lake will only get 25GB/s: quite a bit better than earlier Atoms, which rarely reached 10GB/s no matter what you gave them, dual-channel or not. The biggest benefit is for the iGPU, which probably won’t run Quake well enough, but does fairly well on a 4k desktop at 60Hz: the Chrome 3D Globe View render never ceases to amaze me in terms of what level of 3D interactivity is possible if the code isn’t Microsoft’s Flight Simulator. I haven't really checked, but I'd feel confident to say that it's iGPU performance is similar to the NUC10's Comet Lake UHD, which only has 24EUs but runs at roughly twice the speed. I can't go near the Tiger Lake, but none are meant for gaming.
The NUC has a fan. It’s practically noiseless and even a Prime95/FurMark combo won’t turn it into a howler, but it’s something that can clog and fail eventually. I’d have preferred another fully passive Mini-ITX from ASRock, but that was not to be. Rumor has it, OEMs quite simply refused to take Jasper Lake Atoms from Intel, even at typical contra-revenue bundle deals (effectively for free). Evidently they were stock poison, but I'd love to hear from anyone who knows more.
The Intel NUC is quite extraordinary in that it sports a front USB2 pin-grid connector covered by a rubber cap, that hints a very large signage customer OEM deal as the sole reason for its existence.
It certainly should have been given a 2.5Gbit/s Ethernet port (likely $1 extra cost), but then 1GB Ethernet simply should no longer be sold anywhere today. It also lacks SATA support, which is a bit of a bummer for someone like me, who still has quite a few SATA SSDs on hand, but a very logical choice otherwise.
But note that only two PCIe lanes are available for NVMe use, at least they are 3.0 this time, where all earlier Atoms topped out with 2.0 speeds on PCIe. The two rear USB3 ports are supposed to support 10Gbit/s, not quite Thunderbolt but please consider the TDP levels available.
The NUC chassis otherwise is extremely impressive for this price range, it feels extremely solid and built to last the 10 years perhaps expected from a digital signage device.
It seems a rather competent micro-server device and I'd thoroughly recommend it, if it was a live product. Too bad it’s dead, all you see is left-overs and there is no successor in sight.
Jasper Lake stock is being pushed out... at prices that you may actually want to buy it at. Actually, the older Gemini Lake Refresh is still being sold, in sub-$100 N4020 laptops for example.
The RAM thing is weird. I recall some of the ARK pages lying about the amount of RAM supported on certain Atom models. Like saying 8 GB maximum instead of 32 GB.
The rumor mill has pointed to Alder Lake-N being the upcoming successor to Jasper Lake, with the headline change being a doubling to 8 cores. Presumably, dual-core would be gone for good and the cheapest models would become quad-cores.
There's also the embedded/enterprise-focused Elkhart Lake counterpart to Jasper Lake.
BTW, Elkhart Lake supports in-band ECC, which means it doesn't limit the customer in their selection of RAM. However, it does come at a slight cost in performance and memory capacity.
Too bad all the Elkhart Lake boards I see are rather pricey.
Getting back to ECC, you'd have to move to Atom C-series or P-series to get a SoC with these cores and full ECC support. See links in my post, above.
Are you aware of any boards / PCs with Elkhart Lake that supports in-band ECC? Vendors I talk to seem to indicate that there is some other feature X that gets disabled if you do in-band.. and that feature X is more important for their target market compared to in-band ECC. So, they do not enable in-band ECC in their products even if the processor supports it.
> Are you aware of any boards / PCs with Elkhart Lake that supports in-band ECC?
Sorry, I've not seriously investigated the matter.
> Vendors I talk to seem to indicate that there is some other feature X > that gets disabled if you do in-band.
Wow. I'd love to know more! I figured the main tradeoff was just one of performance (and probably a less significant hit on memory capacity). I wonder why they don't just make it a user-configurable option.
TBH, I don't know specifics about how Intel implements it. I *assume* they simply set aside a chunk of physical address space to hold the ECC bits for the rest of the address space, but that's just a guess.
I had one of those ASRock Apollo Lake boards, but never got it to work. It's possible the RAM I got was incompatible, but it was decent quality (Crucial, IIRC) and their website claimed it worked with that board.
I was sad to see ASRock has no Jasper Lake-based successor, but TBH I'd rather have Elkhart Lake and its in-band ECC-support. I'm just now noticing that Asrock Industrial has some tasty looking options, there. Now, if I can just figure out where to buy a IMB-1003D...
NUCs can take quite a while for the initial boot, even the Core based models.
If I hadn’t been distracted at the time, I’d have already given up on the Jasper Lake NUC working with 64GB: I had been ready to turn it off by the time it showed the logo! Must have lasted something like 30 seconds or so, just to test and tune the RAM, which was DDR4-3200 after all and not quite the DDR4-2900 specs it officially prefers.
Actually, I really hate that vendors increasingly just program at most 2 settings into DIMMs these days, so you can’t recycle them on a different machine.
Once you realized that bits can rot, it’s very difficult to forego ECC. The very first IBM-PCs had parity and I’m not sure when it got dropped from mainline. Once I started running PCs as home servers, I’ve tried to make sure they had ECC memory. My workstations are also all 128GB ECC.
I bought the Atoms mostly to run QA for oVirt, not as a “production” platform: low cost and low power was key, ECC simply not an economically viable option.
They have been running non-stop for years now, with a collective 128GB of RAM and no glitch that I have ever noticed...
The first time I ever heard of inline ECC was in one of your posts here. After a short moment of “bug-eyed disbelief” it seemed to make sense in an era, when little ever happens in RAM below the granularity of a cache line: the days of truly random RAM where all accesses were equally …slow are long past us, I believe the original Compaq 386 was the first to exploit static column RAM.
I believe RAM compression was also implemented by an IBM server chipset many years ago, memory encryption is available on every modern laptop, so inline ECC seems very believable and not extremely costly: I’d just love to have the choice!
As a matter of fact, this gets me asking: Core chips seem to employ ECC practically everywhere on internal registers, caches and data paths, but do Atoms do likewise? I’d guess they would have to for the server variants, so leaving that out for the entry level chips seems almost extra effort, yet I can’t recall hearing any mention one way or another.
I’ve been trying to buy an Alder Lake replacement for a Haswell Xeon server with ASRocks IMB-X1712 mainboard mentioned here that supports DDR4-3200 ECC RAM. Unfortunately that’s another phantom product that never seems available for sale.
> Must have lasted something like 30 seconds or so, just to test and tune the RAM
Does the BIOS have an option to disable it, or at least a "fast boot" option?
> Once you realized that bits can rot, it’s very difficult to forego ECC.
The places where you really want ECC are those where a memory error can get persisted in data of non-trivial value. On fileservers and database servers, it's a must (unless the data is virtually disposable or they're simply providing read-only access).
In the worst case, a memory error can actually cause filesystem corruption. It's unlikely, but the thing to remember about memory errors is that they're not entirely random or isolated. A DRAM chip could conceivably fail in a way that suddenly results a large number of memory errors. This will usually crash the machine (if not using ECC), but you could plausibly suffer data corruption just before that happens.
> I bought the Atoms mostly to run QA for oVirt
My ASRock board was meant to replace my Raspberry Pi as a streaming media server, for in-home use.
> The first time I ever heard of inline ECC was in one of your posts here.
I'm pretty sure the first I'd heard of it was on here, as well. I had a similar reaction as yours, but the more I thought about it, the more sense it made. It'd be hard for me to prefer it when I could have the real deal, but not a bad compromise on something like an Atom-tier platform.
> little ever happens in RAM below the granularity of a cache line
Yeah, you could implement it by blocking off 1/8th of RAM (in truth, you'd only need 1/9th, but 1/8th would keep things aligned more nicely) and associating 8 bytes of ECC information per 64-bytes of physical address space. Depending how you implement it, the hit to memory bandwidth could be as little as 11%, for linear accesses.
TBH, I'm a little more mystified by the concept of memory compression. I guess it'd have to be block based, perhaps decompressing whole pages at a time? Then, when you page fault, some kind of index tells you where the page starts. There'd no doubt be some padding or unused space between the pages (or whatever granularity the blocks are). Perhaps the more interesting aspect would be deciding where to write newly-compressed pages.
Of the three, memory encryption seems the most straight-forward. You would likely have a 1:1 mapping, so the only tricky part is one of key management.
> Unfortunately that’s another phantom product that never seems available for sale.
I'll bet availability is being hampered by just a couple key components being extremely hard to source. I heard some motherboard vendors have been unable to source certain Ethernet MACs. Another example I've heard is RAID controllers.
>> Must have lasted something like 30 seconds or so, just to test and tune the RAM
>Does the BIOS have an option to disable it, or at least a "fast boot" option?
That was only ever an issue for the initial boot with that RAM. Once it has figured out the RAM speed settings any normal boot is at reasonable speeds.
I've research the Elkhart Lake Atoms a bit and they seem quite hard to find. Embedded systems with them sell for eye watering prices.
ZFS was always the typical example for why bit flips could have catastrophical consequences when you cache aggressively and keep key data structure in RAM for months or longer.
I use GlusterFS with VDO de-dup and compression on the Atoms, where a single bit flip could have similarly drastic consequences, but so far I've noticed no issue.
It seems that getting a low power ECC platform is intentionally made difficult, closest I've recently got was with Ryzen 5750G APUs, which isn't that low power nor that cheap.
DDR5 with real ECC seems even worse which is why the ASRock board with the W680 chipset and DDR4 support seems so attractive... and unavailable!
RAM compression: It definitely requires OS support, but other than that seems not too difficult to do. I saw a demo booth at the HiPEAC 2020 conference in Bologna from a Swedish startup I believe, that tries to sell the IP e.g. for integration in RISC-V.
However, that would seem to require the PD10EHI-X6413E model, which is *not* so readily available.
> ASRock board with the W680 chipset
Yeah, I was starting to browse for W680 boards, recently. I wish I could find an ATX (or micro-ATX) with 2x DDR5 slots, but every one I've found is either DDR4 or 4x DDR5 slots. Anyway, I'm not really in a hurry.
> RAM compression: ... seems not too difficult to do. > ...a Swedish startup ... that tries to sell the IP
Okay, think about that for a second. Someone thought it offered enough value and is sufficiently hard that they started a company around it!
BTW, I'm seeing Supermicro (DDR5) W680 boards now in stock! Unfortunately, I can't say the same for DDR5-4800 Unbuffered ECC DIMMs. It's good that I'm in no hurry.
'After suspecting thermal throttling due to the lack of convective cooling for the thermal module, we decided to repeat all the benchmarking for the bare board - i.e, with the top of the chassis removed and the DMICs / WLAN functionality discarded.'
What is this nonsense?
You test the product the way it is designed and sold. If it's designed by idiots it deserves the review it gets.
That is pretty much what I wanted to do, but I ran the numbers just to prove to ECS that it is their chassis that is the problem, and not the internal thermal module. (Typically, we have a back-and-forth with the vendor in case things aren't performing as expected, just to make sure they can reproduce the issue on their side too).
So, once I had the results, I decided to just include it in the review to show readers that the problems lie with the case.
> I ran the numbers just to prove to ECS that it is their chassis that is the problem
And thank you for doing it. Testing as-sold is fine, and those numbers should be the main ones reported. However, supplemental testing to investigate suspected design flaws is the mark of a quality reviewer.
‘As a result, in lieu of CPU overclocking, the biggest thing a user can do to influence higher performance with the Ryzen 7 5800X3D is to use faster DDR4 memory with lower latencies, such as a good DDR4-3600 kit. These settings are also the known sweet spot for AMD's Infinity Fabric Interconnect as set out by AMD.’
> LIVA Z3 tested without its case and one of its main parts.
Yes, but only supplementing their testing in its baseline "as-sold" config. You really ought to go back and re-read stuff before rage-posting about it.
"The numbers for this configuration are referenced using ECS JSLM-MINI (the motherboard's model name) in the relevant sections."
Where data from this configuration is posted, you'll see *both* ECS LIVA Z3 *and* ECS JSLM-MINI, showing just how much the Z3 is hurt by the poor airflow of its chassis.
No, you're flat-out wrong. Trying to shift the subject back on me won't change that. I'd almost feel bad for you embarrassing yourself like this, but you never seem to tire of reminding us what an unpleasant person you are.
FYI: the correct response, when you've been caught leveling a false allegation is: "Oops. Sorry." Anything else just makes it worse.
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
52 Comments
Back to Article
xol - Friday, July 8, 2022 - link
Correction (?)Neither of these reviewed products has a Intel UHD Graphics 605 .. (that's a 14nm Gemini part with 18 EU eg here https://ark.intel.com/content/www/us/en/ark/produc...
.. Intel seems to have not publisher a 'number' for this iGPU and seems to distinguish them by number of EU eg Jasper Lake 24EU eg https://www.intel.co.uk/content/www/uk/en/products...
xol - Friday, July 8, 2022 - link
Somehow messed up the link :UHD 605 https://ark.intel.com/content/www/us/en/ark/produc...
mode_13h - Friday, July 8, 2022 - link
Thanks for your coverage of fanless mini-PCs. However, I really wish you'd include something with "big cores", so we can get a sense of the scale of performance difference between them and Tremont.Another nice-to-have would be at least a few benchmarks including a Raspberry Pi 4. However, it has serious thermal throttling issues, unless it's actively cooled or you use a substantial passive cooling solution.
mode_13h - Friday, July 8, 2022 - link
I guess the ideal comparison would be a Tiger Lake-based system, since that's the same vintage and similar manufacturing tech as Tremont. Probably much harder to find in a fanless mini-PC, unless we're talking about an industrial PC, but I'd love even to see a comparison between two NUCs: Tiger Lake vs. Tremont.mode_13h - Friday, July 8, 2022 - link
Or maybe Ice Lake would be even better, but did they make Ice Lake-based NUCs?abufrejoval - Thursday, July 14, 2022 - link
Yes, Tiger Lake NUCs were made, but also very hard to come by: I have both.In a way they are perfect to showcase the benefit of E/P cores …in the case of Intel: AMD is really another story.
The two NUCs look nearly identical on the outside, but inside they are very different beasts.
For starters: The Tiger Lake NUC11 (i7-1165G7 with 96EU Xe iGPU) is configured with a 64 Watt PL2, a rather long TAU and even the PL2 is 30 Watts by default, I believe. There is a reason it comes with a 90 Watts power brick! I changed PL2 to 50, TAU to 10 seconds and PL1 to 15 Watts to ensure the fan would never howl they way it does with the defaults.
I’ve seen HWinfo report a 5GHz maximum clock, but 4.7GHz is the official top speed. It’s at 64 Watts and near 5GHz clocks that I have measured 1707/5808 Geekbench 4 results on Linux (always a bit faster than on Windows). Jasper Lake doesn’t quite play in the same league at 781/2540 using 3.3 GHz and 10 Watts. In Watts/compute power Tiger Lake looks rather worse than Jasper Lake, but when it comes to rendering a complex web page or recalculating a giant Excel sheet, its sprinting power certainly has it appear much faster.
At 64 Watts the Tiger Lake is a desktop CPU, shoehorned into mobile power envelopes. And when it’s constrained to the levels that passive cooling can manage (see the Supermicro SYS-E100-12T-H review here), it really struggles to deliver that performance. The great thing about the Tiger Lake NUC is that you can change PL1, PL2 and TAU to pretty much anything you want and when you set it to the 10 Watts the Jasper Lake gets to use as an absolute maximum, it starts to do rather badly.
Some of that is because the iGPU always gets preference, leaving close to nothing to the CPU. But some of that is that the remaining power budget forces very low frequencies, where the big Core CPU loses against the Atom cores running at a full speed with these Watts.
Jasper Lake, like all the other Atoms since the J1900, never slows down. I’ve never seen it drop below its “Turbo” clock unless idle, even on a mix of Prime95 and Furmark, and I’ve never seen it exceed 10 Watts of combined CPU+GPU power consumption either.
I also have two Ryzen 5800U based notebooks (1443/7855 on Geekbench4), one of which can be switched between 15 and 28 Watts of TDP. When Tiger Lake and Zen 3 are strictly set to the same power levels, Tiger Lake has to run much slower even with half the cores: Ryzen beats it with a much smaller energy footprint per core. But with Tiger Lake left at the default NUC settings (which a battery powered notebook could not support), its four cores will beat an eight core Zen 3 at 15 Watts in Geekbench, which luckily never seems to exceed TAU.
Intel needs E/P because P cores need too much power at the clock rates they require to beat a Ryzen core, and only with E cores they can hit the efficiency of Zen cores in fully multi-threaded loads.
mode_13h - Thursday, July 14, 2022 - link
Wow, another awesome post! Thanks for taking the time to relate your findings. Very interesting!> the iGPU always gets preference, leaving close to nothing to the CPU.
Very key point, but also one that Intel could conceivably address, to some extent, in future BIOS updates. Not that they're likely to, if it had been on the market for a while when you tested, but it's conceivable.
> in Geekbench, which luckily never seems to exceed TAU.
Another great point! I have never run Geekbench myself, and I haven't noticed reviewers mention this key detail.
Foeketijn - Saturday, September 3, 2022 - link
Don't you want to write for Anand?stanleyipkiss - Friday, July 8, 2022 - link
Zotac makes a fanless zbox with a 1165G7xol - Friday, July 8, 2022 - link
Benches I've seen suggest both are very similar in multi to a i3 low power Skylake eg a ie-6100T (2core 4 thread very common thin client chip) - the gfx capability also seems also a close match for the 24EU part [probably a very similar part with improved HEVC support] (the 32EU N6000 should be better)For single threaded the old Skylake is ~+50% faster., and from Skylake to Alder Lake it's nearly 2x , so nearly 3x from N5100 to i5-12500 for single thread
I have an old fanless Atom Z3735F (22nm) and these new SoCs are a impressive step up (~7x both cpu and gpu) -- I think the Pi Model B latest is very roughly 2x better than that nut no where near the 5100T in any metric.
tldr both benches would have been a wash one way of the other.
mode_13h - Saturday, July 9, 2022 - link
> tldr both benches would have been a wash one way of the other.Huh? If old Skylake is 50% faster, and Jasper Lake is 3.5x as fast as Pi 4 Model B (which seems rather generous), then it wouldn't be "a wash", which is defined as:
13. an action or situation in which the gains and losses are
equal, or closely compensate each other.
(source: http://dict.org/bin/Dict?Form=Dict2&Database=g... )
or
8: any enterprise in which losses and gains cancel out; "at the
end of the year the accounting department showed that it was
a wash"
(source: http://dict.org/bin/Dict?Form=Dict2&Database=w... )
Since both comparisons are projected to be substantially lopsided, I think what you meant to call it is a "washout"?
abufrejoval - Thursday, July 14, 2022 - link
I have a PI4 with 8GB of RAM in a metal case that supports a 2GHz overclock without active cooling: pretty much the best PI you can have these days.I also have an Nvidia Tegra based Jetson Nano with 4GB of RAM.
At 2GHz the PI reaches 272/648 on Geekbench 4, the Tegra has to make do with 206/718 at 1.4GHz. The N6005 Jasper Lake reaches 781/2540 very similar to a Sandy Bridge i7-2600 at 3.8GHz Turbo.
The Jetson Nano actually does reasonably well on my 43" 4k desktop for basic 2D work, because it has a GPU with 128 Maxwell cores. Of course its CPU power is at the level of a Snapdragon 800 mobile phone.
The PI struggles badly at 4k, because the GPU has much less muscle. The slightly faster CPU is hard to notice.
Actually it was when Tom's hardware did a report on a PI compute cluster, that I wanted to retort just how stupid that project was, because you could get a single Jasper Lake Atoms for much less money, that would run rings around that cluster and could in fact simulate it all in software via VMs.
And that's when I found that finally a Jasper Lake NUC was available for purchase at €200 (including VAT) and immediately ordered one of the first and last ever sold here.
And yes, it runs rings around both with roughly 4x the CPU power, 64GB of RAM expandability and quite a reasonable GPU performance on a 4k display.
My favorite usability test is to use the "3D Globe View" on Google Maps under a Chrome based browser on Windows and to then tilt and turn a city landscape there. It's about the most efficient 3D graphics pipeline I've ever seen (puts Flight Simulator to total shame!) and performs quite reasonable on such a Jasper Lake NUC. With Firefox it's much worse on these low power devices, but with a beefy PC you'd never notice.
After quite a bit of tweaking I managed to get it to work on both the PI and the Tegra at 1920x1080 and the Tegra even gave a bit of interactivity thanks to its much stronger GPU. But on the PI that was about one frame a minute.
The PI and Nano are toys and ok for the €100 I spent on each.
A Jasper Lake NUC is quite a reasonable desktop machine and even an interesting micro server for some real workloads.
At €200 (without RAM or storage) the price/performance ratio is very hard to beat, but evidently none of the vendors really want you to know or buy that. I think it's the major reason you never could.
mode_13h - Thursday, July 14, 2022 - link
> At 2GHz the PI reaches 272/648 on Geekbench 4, the Tegra ... 206/718 at 1.4GHz.Keep in mind that Jetson Nano has ostensibly 2x the memory bandwidth of the Pi v4. That surely helps offset the difference in raw CPU performance, as well as with 4k display performance.
Oh, and if that test was with the machines driving a 4k display, then merely refreshing your monitor will have been using a non-insignificant amount of the Pi's memory bandwidth (about 1 GB/s).
> N6005 Jasper Lake reaches 781/2540
Wow! Dual-channel memory configuration, I presume?
> on the PI that was about one frame a minute.
Uh... that sure sounds like you were using a software rendering path. The Pi's GPU is trash, but that's simply atrocious!
> evidently none of the vendors really want you to know or buy that.
> I think it's the major reason you never could.
I'm reasonably confident it's actually just supply chain-related. Intel has been steering its limited fab capacity towards more profitable models and probably steering its limited supply of Jasper Lakes to chromebooks, where they're probably desperate not to lose market share.
timecop1818 - Friday, July 8, 2022 - link
There are Chinese mini PCs withIntel Celeron N5100 that are like 250$ with 16G ram and 256gb sata SSD.
https://www.lazada.com.my/products/walkfish-m6-11t...
there's like 5 different "brands" selling same thing on AliExpress etc. it runs win 10 just fine and is enough for 1080p Minecraft and basic office computing. great deal. most models have Intel 2.5G Ethernet too.
Jorgp2 - Friday, July 8, 2022 - link
I just want a Jasper lake motherboard with plenty of sata and a PCI-E slotmode_13h - Friday, July 8, 2022 - link
You could get SATA, but not PCIe. According to this, Jasper Lake and Elkhart Lake have only x8 PCIe 3.0 and x2 SATA ports.Most boards are probably going to give you a x4 NVMe slot. Then, they could use a 3rd Party SATA controller to give you 4 more ports. Then, if they compromise on the bandwidth to that SATA controller, you can have a second Ethernet port and then a x1 PCIe slot that just might be open-ended (but probably not), to support a graphics card.
Sorry, but they really kneecapped this platform relative to what it could've been. You might do better with some equivalent Atom-branded CPUs. Atom C-series (Parker Ridge) has 16 integrated SATA ports, x32 PCIe 3.0 lanes, and up to 8 cores. P-series (Snow Ridge) has the same, but up to 24 cores.
* https://ark.intel.com/content/www/us/en/ark/produc...
* https://ark.intel.com/content/www/us/en/ark/produc...
mode_13h - Friday, July 8, 2022 - link
Oops, forgot the link for Jasper Lake. For good measure, here's Elkhart Lake, as well.* https://ark.intel.com/content/www/us/en/ark/produc...
* https://ark.intel.com/content/www/us/en/ark/produc...
Thala - Friday, July 8, 2022 - link
Interestingly my 3 years old Surface Pro X scores higher than any of the tested devices in Cinebench R23 under x64 emulation!mode_13h - Saturday, July 9, 2022 - link
The Surface Pro X from 2019 has a Microsoft SQ1 SoC, which is basically a Snapdragon 8cx and consists of 4x Kryo 495 Gold @ 3 GHz+ 4x Kryo 495 Silver @ 1.80 GHz (manufactured on TSMC 7 nm). According to wikichip, these are tweaked A76 and A55 cores. So, that seems credible, if not exactly an outcome I'd have presumed.Something to keep in mind is that Jasper Lake is meant cheap chromebooks. Like, sub-$200 cheap, whereas Snapdragon 8cx is a premium part.
nandnandnand - Saturday, July 9, 2022 - link
They wanted it to be thought of as premium, it's more of an expensive joke. Like Lakefield but with no excuses.https://semiaccurate.com/2021/12/01/qualcomm-8cx-g...
https://www.gizchina.com/2022/01/04/qualcomm-blame...
Snapdragon 7c (Gen 1?) should be more comparable in price to Jasper Lake. I think I've seen that as low as $170-200. Also, the Apcsilmic Dot 1 and ECS LIVA Mini Box QC710 mini PCs recently launched with the 7c starting at around $219.
If the leaks about Alder Lake-N are true, it will shake things up, if the price is right.
IntelUser2000 - Saturday, July 9, 2022 - link
I hope Alderlake-N gets a SoC variant of the 10nm process unlike the HP process for regular Alderlake.nandnandnand - Saturday, July 9, 2022 - link
It will be a shrink from 10nm to Intel 7 (formerly 10nm Enhanced SuperFin), and they should have different cell libraries available. I think they'd want the densest one. I'm sure whatever it becomes will be impressive compared to Jasper Lake.More importantly, I hope we see a lot of the 8-cores.
Thala - Tuesday, July 12, 2022 - link
Of course 8cx devices are more premium. But I was not referring to price, but to what it is technically possible within a 7W power envelope if you include devices, which does not contain Intel CPUs. It just shows that Intel CPUs are incredibly power inefficient independent of price. This includes Lakefield, Jasperlake or whatever CPU Intel designed for the sub 9W TDP market.IntelUser2000 - Saturday, July 9, 2022 - link
In MT yes. It's still quite a bit faster in ST.In ST if not under emulation both CPUs are roughly comparable.
mode_13h - Sunday, July 10, 2022 - link
Ganesh & Ryan:Above, we were talking about how Jasper Lake compares with Skylake and ARM A76 cores. It would be fantastic to have some hard data on this. Why not run SPEC bench on these mini PCs? It can't be any less weird than running it on a iPhone, right?
mode_13h - Sunday, July 10, 2022 - link
Oops, meant to say: "... can't be any weirder than running it on an iPhone"PeachNCream - Sunday, July 10, 2022 - link
A few typos here and there but a good review of hardware that might, for once, realistically land on someone's desk these days. Pity about the Liva's cooling situation. I wonder if chopping open a fair amount of the case top and replacing it with a fan grille or an epoxy-attached bit of non-metallic screen would be sufficient to mitigate the cooling problems while still retaining the bulk of the case itself. Too bad ECS didn't do potential buyers any justice with this design since the cost is otherwise okay.abufrejoval - Monday, July 11, 2022 - link
I can’t help but wonder if Anandtech feels a certain affinity to near-death technology these days…Jasper Lake is unfortunately dead. Whatever is being sold, was manufactured long ago and these systems are the typical leftover stock that Intel is pushing out the back door, as it reduces internal stockpiles to cover long-term warranties: nearly rock bottom, in other words.
I’ve been using Atoms since the J1900 on ASRock Mini-ITX boards, that were 100% passive, simply because they were 100% passive and thus zero noise: a somewhat novel experience ever since I switched from an Apple ][ [clone] to an IBM-AT [clone], which unfortunately started the trend on noisily moving parts.
And one constant has been, that Atoms have always supported way more RAM than advertised. It was 16GB or DDR3 on the DDR3 Atoms, 32GB for the J5005 devices I’ve been running as a oVirt(RHEV) cluster for a couple of years now.
A couple of weeks ago a Jasper Lake or Atlas Canyon NUC11 (NUC11ATKPE) popped up on my radar and since I’d been on a constant lookout for them, I immediately grabbed one, only to see that it might have been both, the first and last gasp of a phantom device: it’s been unavailable since, just as it was for it’s entire official life-time before.
And after putting it through its paces that is quite sad, because it really is a rather strappy and solid little machine, especially at a 2x Raspberry PI4 price point. It performs near identical to a Sandy Bridge i7-2600 and beats the Pi into a pulp with only a tiny helping of extra Watts: 10 vs 28nm does make a difference!
And again, just as you noticed, Jasper Lake will run just fine with far more RAM than Intel wants you to know. That is a constant with every NUC I’ve owned, i7 based NUC8/10/11 all run with 64GB just fine, even if only the Tiger Lake, from which I borrowed the dual 32GB DDR4-3200 SO-DIMMs for the Jasper Lake evaluation, did so officially.
But where my Tiger Lake will squeeze 40GB/s from that RAM, Jasper Lake will only get 25GB/s: quite a bit better than earlier Atoms, which rarely reached 10GB/s no matter what you gave them, dual-channel or not. The biggest benefit is for the iGPU, which probably won’t run Quake well enough, but does fairly well on a 4k desktop at 60Hz: the Chrome 3D Globe View render never ceases to amaze me in terms of what level of 3D interactivity is possible if the code isn’t Microsoft’s Flight Simulator. I haven't really checked, but I'd feel confident to say that it's iGPU performance is similar to the NUC10's Comet Lake UHD, which only has 24EUs but runs at roughly twice the speed. I can't go near the Tiger Lake, but none are meant for gaming.
The NUC has a fan. It’s practically noiseless and even a Prime95/FurMark combo won’t turn it into a howler, but it’s something that can clog and fail eventually. I’d have preferred another fully passive Mini-ITX from ASRock, but that was not to be. Rumor has it, OEMs quite simply refused to take Jasper Lake Atoms from Intel, even at typical contra-revenue bundle deals (effectively for free). Evidently they were stock poison, but I'd love to hear from anyone who knows more.
The Intel NUC is quite extraordinary in that it sports a front USB2 pin-grid connector covered by a rubber cap, that hints a very large signage customer OEM deal as the sole reason for its existence.
It certainly should have been given a 2.5Gbit/s Ethernet port (likely $1 extra cost), but then 1GB Ethernet simply should no longer be sold anywhere today. It also lacks SATA support, which is a bit of a bummer for someone like me, who still has quite a few SATA SSDs on hand, but a very logical choice otherwise.
But note that only two PCIe lanes are available for NVMe use, at least they are 3.0 this time, where all earlier Atoms topped out with 2.0 speeds on PCIe. The two rear USB3 ports are supposed to support 10Gbit/s, not quite Thunderbolt but please consider the TDP levels available.
The NUC chassis otherwise is extremely impressive for this price range, it feels extremely solid and built to last the 10 years perhaps expected from a digital signage device.
It seems a rather competent micro-server device and I'd thoroughly recommend it, if it was a live product. Too bad it’s dead, all you see is left-overs and there is no successor in sight.
nandnandnand - Tuesday, July 12, 2022 - link
Jasper Lake stock is being pushed out... at prices that you may actually want to buy it at. Actually, the older Gemini Lake Refresh is still being sold, in sub-$100 N4020 laptops for example.The RAM thing is weird. I recall some of the ARK pages lying about the amount of RAM supported on certain Atom models. Like saying 8 GB maximum instead of 32 GB.
The rumor mill has pointed to Alder Lake-N being the upcoming successor to Jasper Lake, with the headline change being a doubling to 8 cores. Presumably, dual-core would be gone for good and the cheapest models would become quad-cores.
There's also the embedded/enterprise-focused Elkhart Lake counterpart to Jasper Lake.
mode_13h - Wednesday, July 13, 2022 - link
BTW, Elkhart Lake supports in-band ECC, which means it doesn't limit the customer in their selection of RAM. However, it does come at a slight cost in performance and memory capacity.Too bad all the Elkhart Lake boards I see are rather pricey.
Getting back to ECC, you'd have to move to Atom C-series or P-series to get a SoC with these cores and full ECC support. See links in my post, above.
ganeshts - Wednesday, July 13, 2022 - link
Are you aware of any boards / PCs with Elkhart Lake that supports in-band ECC? Vendors I talk to seem to indicate that there is some other feature X that gets disabled if you do in-band.. and that feature X is more important for their target market compared to in-band ECC. So, they do not enable in-band ECC in their products even if the processor supports it.mode_13h - Thursday, July 14, 2022 - link
> Are you aware of any boards / PCs with Elkhart Lake that supports in-band ECC?Sorry, I've not seriously investigated the matter.
> Vendors I talk to seem to indicate that there is some other feature X
> that gets disabled if you do in-band.
Wow. I'd love to know more! I figured the main tradeoff was just one of performance (and probably a less significant hit on memory capacity). I wonder why they don't just make it a user-configurable option.
TBH, I don't know specifics about how Intel implements it. I *assume* they simply set aside a chunk of physical address space to hold the ECC bits for the rest of the address space, but that's just a guess.
ganeshts - Tuesday, July 12, 2022 - link
Looks like you will leave me with nothing to write about for the Atlas Canyon review coming up later this week :)abufrejoval - Thursday, July 14, 2022 - link
If you keep me updated on the things in your pipeline, I'll make sure not to spoil things ;-)mode_13h - Thursday, July 14, 2022 - link
But I love your posts! I'll bet < 1% of the article readers look this deep into the comments.mode_13h - Wednesday, July 13, 2022 - link
Nice review!I had one of those ASRock Apollo Lake boards, but never got it to work. It's possible the RAM I got was incompatible, but it was decent quality (Crucial, IIRC) and their website claimed it worked with that board.
I was sad to see ASRock has no Jasper Lake-based successor, but TBH I'd rather have Elkhart Lake and its in-band ECC-support. I'm just now noticing that Asrock Industrial has some tasty looking options, there. Now, if I can just figure out where to buy a IMB-1003D...
abufrejoval - Thursday, July 14, 2022 - link
NUCs can take quite a while for the initial boot, even the Core based models.If I hadn’t been distracted at the time, I’d have already given up on the Jasper Lake NUC working with 64GB: I had been ready to turn it off by the time it showed the logo! Must have lasted something like 30 seconds or so, just to test and tune the RAM, which was DDR4-3200 after all and not quite the DDR4-2900 specs it officially prefers.
Actually, I really hate that vendors increasingly just program at most 2 settings into DIMMs these days, so you can’t recycle them on a different machine.
Once you realized that bits can rot, it’s very difficult to forego ECC. The very first IBM-PCs had parity and I’m not sure when it got dropped from mainline. Once I started running PCs as home servers, I’ve tried to make sure they had ECC memory. My workstations are also all 128GB ECC.
I bought the Atoms mostly to run QA for oVirt, not as a “production” platform: low cost and low power was key, ECC simply not an economically viable option.
They have been running non-stop for years now, with a collective 128GB of RAM and no glitch that I have ever noticed...
The first time I ever heard of inline ECC was in one of your posts here. After a short moment of “bug-eyed disbelief” it seemed to make sense in an era, when little ever happens in RAM below the granularity of a cache line: the days of truly random RAM where all accesses were equally …slow are long past us, I believe the original Compaq 386 was the first to exploit static column RAM.
I believe RAM compression was also implemented by an IBM server chipset many years ago, memory encryption is available on every modern laptop, so inline ECC seems very believable and not extremely costly: I’d just love to have the choice!
As a matter of fact, this gets me asking: Core chips seem to employ ECC practically everywhere on internal registers, caches and data paths, but do Atoms do likewise? I’d guess they would have to for the server variants, so leaving that out for the entry level chips seems almost extra effort, yet I can’t recall hearing any mention one way or another.
I’ve been trying to buy an Alder Lake replacement for a Haswell Xeon server with ASRocks IMB-X1712 mainboard mentioned here that supports DDR4-3200 ECC RAM. Unfortunately that’s another phantom product that never seems available for sale.
mode_13h - Thursday, July 14, 2022 - link
> Must have lasted something like 30 seconds or so, just to test and tune the RAMDoes the BIOS have an option to disable it, or at least a "fast boot" option?
> Once you realized that bits can rot, it’s very difficult to forego ECC.
The places where you really want ECC are those where a memory error can get persisted in data of non-trivial value. On fileservers and database servers, it's a must (unless the data is virtually disposable or they're simply providing read-only access).
In the worst case, a memory error can actually cause filesystem corruption. It's unlikely, but the thing to remember about memory errors is that they're not entirely random or isolated. A DRAM chip could conceivably fail in a way that suddenly results a large number of memory errors. This will usually crash the machine (if not using ECC), but you could plausibly suffer data corruption just before that happens.
> I bought the Atoms mostly to run QA for oVirt
My ASRock board was meant to replace my Raspberry Pi as a streaming media server, for in-home use.
> The first time I ever heard of inline ECC was in one of your posts here.
I'm pretty sure the first I'd heard of it was on here, as well. I had a similar reaction as yours, but the more I thought about it, the more sense it made. It'd be hard for me to prefer it when I could have the real deal, but not a bad compromise on something like an Atom-tier platform.
> little ever happens in RAM below the granularity of a cache line
Yeah, you could implement it by blocking off 1/8th of RAM (in truth, you'd only need 1/9th, but 1/8th would keep things aligned more nicely) and associating 8 bytes of ECC information per 64-bytes of physical address space. Depending how you implement it, the hit to memory bandwidth could be as little as 11%, for linear accesses.
TBH, I'm a little more mystified by the concept of memory compression. I guess it'd have to be block based, perhaps decompressing whole pages at a time? Then, when you page fault, some kind of index tells you where the page starts. There'd no doubt be some padding or unused space between the pages (or whatever granularity the blocks are). Perhaps the more interesting aspect would be deciding where to write newly-compressed pages.
Of the three, memory encryption seems the most straight-forward. You would likely have a 1:1 mapping, so the only tricky part is one of key management.
> Unfortunately that’s another phantom product that never seems available for sale.
I'll bet availability is being hampered by just a couple key components being extremely hard to source. I heard some motherboard vendors have been unable to source certain Ethernet MACs. Another example I've heard is RAID controllers.
abufrejoval - Thursday, July 14, 2022 - link
>> Must have lasted something like 30 seconds or so, just to test and tune the RAM>Does the BIOS have an option to disable it, or at least a "fast boot" option?
That was only ever an issue for the initial boot with that RAM. Once it has figured out the RAM speed settings any normal boot is at reasonable speeds.
I've research the Elkhart Lake Atoms a bit and they seem quite hard to find. Embedded systems with them sell for eye watering prices.
ZFS was always the typical example for why bit flips could have catastrophical consequences when you cache aggressively and keep key data structure in RAM for months or longer.
I use GlusterFS with VDO de-dup and compression on the Atoms, where a single bit flip could have similarly drastic consequences, but so far I've noticed no issue.
It seems that getting a low power ECC platform is intentionally made difficult, closest I've recently got was with Ryzen 5750G APUs, which isn't that low power nor that cheap.
DDR5 with real ECC seems even worse which is why the ASRock board with the W680 chipset and DDR4 support seems so attractive... and unavailable!
RAM compression: It definitely requires OS support, but other than that seems not too difficult to do. I saw a demo booth at the HiPEAC 2020 conference in Bologna from a Swedish startup I believe, that tries to sell the IP e.g. for integration in RISC-V.
mode_13h - Friday, July 15, 2022 - link
> Elkhart Lake Atoms a bit and they seem quite hard to find.They exist, if expensive and uncommon: https://www.newegg.com/p/1JW-003Z-00026
According to the manufacturer's site, it even seems to support in-band ECC:
https://www.mitacmct.com/IndustrialMotherboard=PD1...
However, that would seem to require the PD10EHI-X6413E model, which is *not* so readily available.
> ASRock board with the W680 chipset
Yeah, I was starting to browse for W680 boards, recently. I wish I could find an ATX (or micro-ATX) with 2x DDR5 slots, but every one I've found is either DDR4 or 4x DDR5 slots. Anyway, I'm not really in a hurry.
> RAM compression: ... seems not too difficult to do.
> ...a Swedish startup ... that tries to sell the IP
Okay, think about that for a second. Someone thought it offered enough value and is sufficiently hard that they started a company around it!
mode_13h - Tuesday, July 19, 2022 - link
BTW, I'm seeing Supermicro (DDR5) W680 boards now in stock! Unfortunately, I can't say the same for DDR5-4800 Unbuffered ECC DIMMs. It's good that I'm in no hurry.Oxford Guy - Tuesday, July 12, 2022 - link
'After suspecting thermal throttling due to the lack of convective cooling for the thermal module, we decided to repeat all the benchmarking for the bare board - i.e, with the top of the chassis removed and the DMICs / WLAN functionality discarded.'What is this nonsense?
You test the product the way it is designed and sold. If it's designed by idiots it deserves the review it gets.
Oxford Guy - Tuesday, July 12, 2022 - link
ECS knows what ventilation slots are. The first model of Liva has them on the top, bottom, and sides of the unit!ganeshts - Tuesday, July 12, 2022 - link
That is pretty much what I wanted to do, but I ran the numbers just to prove to ECS that it is their chassis that is the problem, and not the internal thermal module. (Typically, we have a back-and-forth with the vendor in case things aren't performing as expected, just to make sure they can reproduce the issue on their side too).So, once I had the results, I decided to just include it in the review to show readers that the problems lie with the case.
mode_13h - Wednesday, July 13, 2022 - link
> I ran the numbers just to prove to ECS that it is their chassis that is the problemAnd thank you for doing it. Testing as-sold is fine, and those numbers should be the main ones reported. However, supplemental testing to investigate suspected design flaws is the mark of a quality reviewer.
Keep up the good work!
Oxford Guy - Saturday, July 16, 2022 - link
Zen 1 and 2, tested with JEDEC slowwww RAM and not also with any XMP profiles.LIVA Z3 tested without its case and one of its main parts.
I guess removing the protective case and the wireless Internet of a computer is more normal ordinary user behavior than turning on XMP in BIOS.
Just trying to understand the big picture here...
Oxford Guy - Saturday, July 16, 2022 - link
And, of course, that continues:‘As a result, in lieu of CPU overclocking, the biggest thing a user can do to influence higher performance with the Ryzen 7 5800X3D is to use faster DDR4 memory with lower latencies, such as a good DDR4-3600 kit. These settings are also the known sweet spot for AMD's Infinity Fabric Interconnect as set out by AMD.’
proceeds to test with 3200-speed RAM...
mode_13h - Sunday, July 17, 2022 - link
> LIVA Z3 tested without its case and one of its main parts.Yes, but only supplementing their testing in its baseline "as-sold" config. You really ought to go back and re-read stuff before rage-posting about it.
"The numbers for this configuration are referenced using ECS JSLM-MINI (the motherboard's model name) in the relevant sections."
Where data from this configuration is posted, you'll see *both* ECS LIVA Z3 *and* ECS JSLM-MINI, showing just how much the Z3 is hurt by the poor airflow of its chassis.
Oxford Guy - Sunday, July 17, 2022 - link
'You really ought to go back and re-read stuff before rage-posting about it.'Your usual ad hom routine won't change this interesting inconsistency.
mode_13h - Sunday, July 17, 2022 - link
> Your usual ad hom routineNo, you're flat-out wrong. Trying to shift the subject back on me won't change that. I'd almost feel bad for you embarrassing yourself like this, but you never seem to tire of reminding us what an unpleasant person you are.
FYI: the correct response, when you've been caught leveling a false allegation is: "Oops. Sorry." Anything else just makes it worse.
Oxford Guy - Monday, July 18, 2022 - link
‘No, you're flat-out wrong’That’s nice.
imofullform - Thursday, July 21, 2022 - link
Nice Post