Looking at the specs and benches, it seems pretty reasonable against the 3975WX. I suspect that's going to be a tougher compare when Milan-based TR Pro launches, though...
Milan TR might be unobtanium. Heck, many Rome TR SKUs (like the 3990X that will inevitably draw some comparisons) are *still* impossible to get at MSRP.
The workstation chips don't have UPI lanes, so AMD would not support multi-socket coherency.
It will be interesting to see if Intel enables multi-socket workstations that use cxl for coherency when Sapphire Rapids workstation chips are launched. I could see this working with their new IPU idea.
Price/perf is not really competitive with TR Pro, and not remotely with normal 32-core TR.
The only compelling reason to buy WX-3300 is for the system-level upgrades (PCIe 4.0 lanes; memory) for those compelled to stay with Intel (or who cannot source TR).
The 16-core seems most compelling, if you needed more I/O than a Ryzen AM4 system could provide and yet 16 cores was enough compute. Multi-GPU deep learning workstations come to mind.
TR Pro has a lot more PCIe lanes, making it very attractive for a workstation build. Hell everything about the TRPro is better. Why would anyone spend MORE money for an inferior product? If you say AVX512, I'm gonna slap you silly...
> Why would anyone spend MORE money for an inferior product?
Some people are stuck in the Intel ecosystem, for whatever reason. Whether it's what their software vendor is certifying or maybe it's what their corporate IT department will purchase. A lot of big OEMs, like Dell, are still mostly in the Intel camp.
Availability is another reason.
> If you say AVX512, I'm gonna slap you silly...
I've criticized it many times, but it really comes down to the software you need to run. If it's CPU-bottlenecked and they've optimized it for AVX-512 (which could be as simple as using Intel-provided optimized libraries), then it's a legitimate consideration. More cores can *usually* offset the difference, but not always.
As you point out, Intel is facing a real challenge in Threadripper. I see this as mostly a stop-gap measure, to keep them marginally in the game until they can release something based on Sapphire Rapids.
Compared to the W-3345 at $2499, the EPYC 7443P-based workstations I’ve just built have twice the PCIe lanes, higher all-core boost clocks, over 3x the L3 cache, use a third less power and cost $1337 per CPU.
With these high-frequency EPYC Milan chips about, TR Pro is a dead platform until the Zen 3 based chips launch, if ever.
I’m in the UK; they are about if you’re willing to put the time in to hunt.
Got 5x units at MSRP from various resellers, and 1x unit from the ‘dark channel’ at ~60% markup, which I was happy to pay as these systems needed to be built.
I found a webiste that virtually contains everything i nnedded. I was browsing on my computer one day and i came to about a website called sheepskinplanet I was curious about what they got to offer so i started exploring their products and to my surprise i found authentic rugs made up of pure Australian sheep skin. They caught my eye in an instance and i ordred up a stack of pelts for my home and my office. Now i am so delighted with their product that i couldn't stop myself from praising their quality and products. I am happy and satisfied customer and i would suggest you to follow in my foot prints and order up. I will leave you with the link. <a href="https://sheepskinplanet.com/"> Medical Sheepskin </a>
"Intel has not disclosed the die configurations of the lower core count parts, however the 28+ core parts are almost certainly using the 40 physical core configurations."
I'm surprised that the W-3345 24 core processor is based on the XCC die instead of a 28 core die. The W-3335 and W-3323 on the other hand could just as well be based on a 18 core die. Is Intel saving all good die for servers where power consumption is more strict?
Although Intel was supposedly working on a 16C LCC ICL-SP die, they have yet to release any products based on it, so it was likely scrapped prior to volume production.
I'm guessing the 40C die responds better to dumping power into it than the 28C when trying to hit those all-core turbo clocks, and you have 43% more cores to choose from when looking for that golden core that can reach the max single-core turbo. As long as Intel can make their margin, it doesn't matter which die they use. And comparing pricing to the 3rd Gen Xeon Scalable SKUs, they certainly seem to be shooting for that.
Cost is typically no object for a workstation - glad to see AT doing Intel’s marketing department’s work for them in cutting the top AMD SKU’s out of the comparison.
That's highly dependent on the use. If the workstation is for running software with licensing costs in the multiple- $10k / year range, then the hardware costs aren't a major consideration.
However, my employer used to buy workstations as the standard (non-laptop) platform for software developers and subsequently switched to mainstream desktops, as workstation costs have climbed.
Also, I built my own socket-2011 workstation that's due for a replacement. I have been priced out of the workstation market, for several years now. Not only CPUs but also motherboards are a lot more expensive than what I'm willing to spend. Fortunately, the capabilities of mainstream desktops have really come up to bridge the gap, mostly thanks to AMD.
Those embarrassingly-parallel batch workloads aren't very interesting. They are effectively benchmarks of how many cores you have. You can predict the results from just looking at the price per core.
They're influenced by the memory & cache subsystem, as well as things like all-core turbo workload.
It's true that compilation benchmarks favor machines with more cores, but when you're comparing two CPUs that are both 32 cores / 64 threads, the outcome is not a foregone conclusion.
More importantly, it's a common workload that's certainly more prevalent among real workstation users than any of the other benchmarks included in the article.
Finally, Ian already *does* compilation benchmarks, in some of the reviews. So, it hopefully shouldn't be a big deal to include them, here. If you find them uninteresting, you're free to disregard them.
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
45 Comments
Back to Article
sara-windowdeco - Thursday, July 29, 2021 - link
This is already a hard sell compared to the old TR let alone the upcoming TR.SarahKerrigan - Thursday, July 29, 2021 - link
Looking at the specs and benches, it seems pretty reasonable against the 3975WX. I suspect that's going to be a tougher compare when Milan-based TR Pro launches, though...brucethemoose - Thursday, July 29, 2021 - link
That all depends on pricing and availability.Milan TR might be unobtanium. Heck, many Rome TR SKUs (like the 3990X that will inevitably draw some comparisons) are *still* impossible to get at MSRP.
JayNor - Thursday, July 29, 2021 - link
The workstation chips don't have UPI lanes, so AMD would not support multi-socket coherency.It will be interesting to see if Intel enables multi-socket workstations that use cxl for coherency when Sapphire Rapids workstation chips are launched. I could see this working with their new IPU idea.
JayNor - Thursday, July 29, 2021 - link
correct ... meant the Intel workstation chips would not support multi-socket coherency...mode_13h - Monday, August 2, 2021 - link
> Looking at the specs and benches, it seems pretty reasonable against the 3975WX.Let's compare frequencies & pricing, and also with non-Pro 32-core TR:
Model - Freq (GHz) - MSRP
---------------------------------------
Xeon W-3365 - 3.5/4.0 - $3500
TR 3975WX - 3.5/4.2 - $2750
TR 3970X - 3.7/4.5 - $2000
Price/perf is not really competitive with TR Pro, and not remotely with normal 32-core TR.
The only compelling reason to buy WX-3300 is for the system-level upgrades (PCIe 4.0 lanes; memory) for those compelled to stay with Intel (or who cannot source TR).
The 16-core seems most compelling, if you needed more I/O than a Ryzen AM4 system could provide and yet 16 cores was enough compute. Multi-GPU deep learning workstations come to mind.
YB1064 - Wednesday, August 4, 2021 - link
TR Pro has a lot more PCIe lanes, making it very attractive for a workstation build. Hell everything about the TRPro is better. Why would anyone spend MORE money for an inferior product? If you say AVX512, I'm gonna slap you silly...mode_13h - Thursday, August 5, 2021 - link
> Why would anyone spend MORE money for an inferior product?Some people are stuck in the Intel ecosystem, for whatever reason. Whether it's what their software vendor is certifying or maybe it's what their corporate IT department will purchase. A lot of big OEMs, like Dell, are still mostly in the Intel camp.
Availability is another reason.
> If you say AVX512, I'm gonna slap you silly...
I've criticized it many times, but it really comes down to the software you need to run. If it's CPU-bottlenecked and they've optimized it for AVX-512 (which could be as simple as using Intel-provided optimized libraries), then it's a legitimate consideration. More cores can *usually* offset the difference, but not always.
As you point out, Intel is facing a real challenge in Threadripper. I see this as mostly a stop-gap measure, to keep them marginally in the game until they can release something based on Sapphire Rapids.
shanti_shakti - Thursday, July 29, 2021 - link
Compared to the W-3345 at $2499, the EPYC 7443P-based workstations I’ve just built have twice the PCIe lanes, higher all-core boost clocks, over 3x the L3 cache, use a third less power and cost $1337 per CPU.With these high-frequency EPYC Milan chips about, TR Pro is a dead platform until the Zen 3 based chips launch, if ever.
Mikewind Dale - Thursday, July 29, 2021 - link
Where did you find an EPYC 7443P to buy? It's easy to find ThreadRipper Pros for sale, but everywhere I saw an EPYC 7443P listed it as backordered.shanti_shakti - Friday, July 30, 2021 - link
I’m in the UK; they are about if you’re willing to put the time in to hunt.Got 5x units at MSRP from various resellers, and 1x unit from the ‘dark channel’ at ~60% markup, which I was happy to pay as these systems needed to be built.
kgardas - Thursday, July 29, 2021 - link
Unless you are riding AVX512 wave. Then whatever Intel produces with it makes more sense I guess.mode_13h - Monday, August 2, 2021 - link
Even then, Intel's $/core ratio is a lot higher than AMD's. Probably also $/perf, on most AVX2/512 workloads.[email protected] - Wednesday, August 4, 2021 - link
I found a webiste that virtually contains everything i nnedded. I was browsing on my computer one day and i came to about a website called sheepskinplanetI was curious about what they got to offer so i started exploring their products and to my surprise i found authentic rugs made up of pure Australian sheep skin.
They caught my eye in an instance and i ordred up a stack of pelts for my home and my office. Now i am so delighted with their product that i couldn't stop myself from
praising their quality and products. I am happy and satisfied customer and i would suggest you to follow in my foot prints and order up. I will leave you with the link.
<a href="https://sheepskinplanet.com/"> Medical Sheepskin </a>
mode_13h - Thursday, August 5, 2021 - link
spammer.bwj - Thursday, July 29, 2021 - link
The W-3200 would have been a more useful baseline.Timoo - Thursday, July 29, 2021 - link
Too much Intel-boasters. Sorry.Just another monolith trying to defeat their eternal loss.
JayNor - Thursday, July 29, 2021 - link
"Although not disclosed, we are under the impression that the Xeon W-3300 parts have dual AVX512 FMAs per core."https://ark.intel.com/content/www/us/en/ark/produc...
product specification does say 2 avx512 FMA units
JayNor - Thursday, July 29, 2021 - link
also shows two avx512 FMA units per core on W-3323 ... the 12 core chip.https://ark.intel.com/content/www/us/en/ark/produc...
repoman27 - Thursday, July 29, 2021 - link
"Intel has not disclosed the die configurations of the lower core count parts, however the 28+ core parts are almost certainly using the 40 physical core configurations."ARK also tells us the dies used:
W-3375 = D2 stepping (40C XCC die)
W-3365 = D2 stepping (40C XCC die)
W-3345 = D2 stepping (40C XCC die)
W-3335 = M1 stepping (28C HCC die)
W-3323 = M1 stepping (28C HCC die)
Rudde - Friday, July 30, 2021 - link
I'm surprised that the W-3345 24 core processor is based on the XCC die instead of a 28 core die. The W-3335 and W-3323 on the other hand could just as well be based on a 18 core die. Is Intel saving all good die for servers where power consumption is more strict?Wrs - Friday, July 30, 2021 - link
Probably a clue that 10nm yield is poor. They can't wait to turn a new chapter on process node quality.Spunjji - Friday, July 30, 2021 - link
Almost certainly. They're clearly not getting a lot of good dies...repoman27 - Friday, July 30, 2021 - link
Although Intel was supposedly working on a 16C LCC ICL-SP die, they have yet to release any products based on it, so it was likely scrapped prior to volume production.I'm guessing the 40C die responds better to dumping power into it than the 28C when trying to hit those all-core turbo clocks, and you have 43% more cores to choose from when looking for that golden core that can reach the max single-core turbo. As long as Intel can make their margin, it doesn't matter which die they use. And comparing pricing to the 3rd Gen Xeon Scalable SKUs, they certainly seem to be shooting for that.
JayNor - Friday, July 30, 2021 - link
you're misquoting arc. w-3323 is 12 cores.Oxford Guy - Thursday, July 29, 2021 - link
'Although not disclosed, we are under the impression that the Xeon W-3300 parts have dual AVX512 FMAs per core.'Bizarre that such basic information hasn't been provided.
Sahrin - Thursday, July 29, 2021 - link
Cost is typically no object for a workstation - glad to see AT doing Intel’s marketing department’s work for them in cutting the top AMD SKU’s out of the comparison.mode_13h - Monday, August 2, 2021 - link
> Cost is typically no object for a workstationThat's highly dependent on the use. If the workstation is for running software with licensing costs in the multiple- $10k / year range, then the hardware costs aren't a major consideration.
However, my employer used to buy workstations as the standard (non-laptop) platform for software developers and subsequently switched to mainstream desktops, as workstation costs have climbed.
Also, I built my own socket-2011 workstation that's due for a replacement. I have been priced out of the workstation market, for several years now. Not only CPUs but also motherboards are a lot more expensive than what I'm willing to spend. Fortunately, the capabilities of mainstream desktops have really come up to bridge the gap, mostly thanks to AMD.
flgt - Thursday, July 29, 2021 - link
Intel, only when AMD is out of stock and you have no other options.JayNor - Friday, July 30, 2021 - link
3323 is 12 corehttps://ark.intel.com/content/www/us/en/ark/produc...
dsplover - Saturday, July 31, 2021 - link
They better make sure Alder Lake isn’t delayed because this is only competitive with other Xeon designs.zamroni - Saturday, July 31, 2021 - link
Ice Lame architecture for Xeon Weakmode_13h - Monday, August 2, 2021 - link
Ian, please include software compilation in your standard battery of workstation tests. Linux kernel or LLVM compilation, for instance.As one datapoint, Linus Torvalds uses a TR 3970X. As another, ESR used a TR workstation to convert GCC's source repository over to git.
bwj - Monday, August 2, 2021 - link
Those embarrassingly-parallel batch workloads aren't very interesting. They are effectively benchmarks of how many cores you have. You can predict the results from just looking at the price per core.mode_13h - Tuesday, August 3, 2021 - link
They're influenced by the memory & cache subsystem, as well as things like all-core turbo workload.It's true that compilation benchmarks favor machines with more cores, but when you're comparing two CPUs that are both 32 cores / 64 threads, the outcome is not a foregone conclusion.
More importantly, it's a common workload that's certainly more prevalent among real workstation users than any of the other benchmarks included in the article.
Finally, Ian already *does* compilation benchmarks, in some of the reviews. So, it hopefully shouldn't be a big deal to include them, here. If you find them uninteresting, you're free to disregard them.
Oxford Guy - Tuesday, August 3, 2021 - link
Interesting isn’t the point. Relevant is.Makste - Thursday, August 5, 2021 - link
Thanks