AMD’s Epyc Pummels Intel’s New Xeon-W Workstation CPUs
semiaccurate.com
semiaccurate.com
It's pretty clear that now AMD is speed competitive, people will look at features and this artificial segmentation will push more people to AMD.
Eg, from the article:
You may notice the Silver and Bronze tiers of Xeons are gone, as are the -M SKUs. This leaves the workstation parts crippled, you can’t get more than 768GB per socket on parts meant for heavy memory intensive tasks like CAD, FEA, and modeling large data sets.... AMD’s Epyc on the other hand can put up to 2TB in a single socket, 4TB with 256GB DIMMs versus Intel’s 756GB, an artificially enforced cap so DIMM side is irrelevant.
Also, the Xeon 8180 is $10,000?! I used to spec servers, and I don't remember seeing CPUs anywhere near that price. Is that common in the workstation market?
Actually, for me following at home ... it seams like they have a whole new segmentation scheme. So not killing it, just changing it.
Also, the $10k one isn't even the most expensive. The Xeon Platinum 8180M is $13,011.
It will definitely be great for competition if AMD can become more competitive with Intel's latest chips.
This is an opinionated editorial by someone who knows the market pretty well, and there is a place for that.
Benchmarks:
- Top AMD Threadripper model vs Top Intel I9 X-series currently being sold (same cpu as Xeon W-2155, just without ECC): https://www.phoronix.com/scan.php?page=article&item=amd-tr-1...
14 and 18 core Xeon-W and i9 processors have only been announced, but are not available yet.
We already know that the Intel 18-core parts will have a base clock of 2.3 GHz vs 3.3 GHz of the one tested above, both parts have 4 memory channels.
The 16-core AMD Threadripper from that benchmark runs at 3.4GHz base clock and has 4 memory channels, the 32-core AMD EPYC runs at 2.2GHz base clock and has 8 memory channels.
-- Educated guesses and extrapolation starts here--
Intel 10-core vs 18-core:
+80% more cores
-30% base clock
Assuming perfect core & clock scaling: 1.8 * 0.7 = 1.26
AMD 16-core vs 32-core
+100% more cores
-35% base clock
Assuming perfect core & clock scaling: 2 * 0.65 = 1.3
Even ignoring double memory bandwidth AMD is likely to increase any advantage (or catch up in AVX2-heavy workloads) over Intel.
I'm only looking at base clocks, because Intel CPUs are not able to maintain boost clock for any meaningful amount of time, even with water cooling - the package doesn't transfer heat from the die fast enough. AMDs fare quite a bit better, but few people will bother with water cooling their workstations.
Ryzen CPUs support ECC if mobo does, but were not formally validated for that.
Threadripper and Epyc both support ECC and have been validated.
Still, EPYC wins on max memory (2TB per processor) and number of PCI-e lanes (128 vs 48). And of course the price.
No, the top Xeon-W processor has 18 cores. The 28 core processor is a Xenon Platinum.
- Xeon-w cpus seem to be i9 HEDT just without disabled ECC - everything else (beside a price premium) is the same
- Epyc CPUs are server CPUs and I'm not aware of any desktop motherboards for them, but considering 32 core Epyc has the same TDP as Threadripper they could release a desktop variant (different socket)
- Xeon-W and i9s with more than 10 cores have only been announced, you can't buy them yet
so it's a comparison of parts you can't get your hands on with either parts in a different space or ones that were not announced yet
Not a particularly impressive improvement since the cost to a dual socket node adds $1,000.
> Intel doesn’t want to be put in the position of admitting they are charging significantly more for a new product that runs about half the speed of their cheaper competition.
Unfortunately, people are locked into the Xeon line - so it doesn't matter a ton for now. In the coming years if AMD can keep up, they'll win - over and over time.
(Currently sitting on two E5-2660v2 for reference)
I think instead of upgrading this thing any more I'll just sell it and build a new EPYC system once the parts are available outside of OEM configurations. At $600 the EPYC 7281 is a steal - 16c/32t, a full 128 PCIe lanes plus four memory channels (with 2TB of memory supported) on a single socket.
I'm still pretty happy with my ~3yo pc, I did ditch windows as my desktop OS earlier this year. (aside: win10's aggressive advertising in the OS finally irritated me too much, and I'm not the only one).
I'm hoping that the Xeon E3 flavor of the i7 is out shortly afterwards. Should make for a great workstation and guarantees ECC.
I was shopping for ECC threadripper boards and they often don't mention ECC. I'm hoping AMD fixes that and brings out a threadripper down to $350 (with the Intel x-series starts).
The math doesn't work out here: 32 cores isn't 14 cores more than 28 cores. I think this was a typo and the author meant to compare with the 18 core Skylake-X, which exists (whereas there is no 28 core model).
You really need to hate your hard earned $ with a passion to buy such rubbish with an Intel logo on it.
He was also one of the first reporters to report Intel's secret paymenst to Dell to not use AMD Athlon CPUs, that ultimately greatly limited AMD's ability to generate revenue to increase its production capacity at its fab.
Believe it or not, Intel acted like a mini-Microsoft back in AMD's Athlon heydays, and that pissed off a lot of people in the industry.
Charlie holds many grudges; he pissed NVIDIA off enough that they denied him access after which his been dumping dirt on them whenever he can, justified or not.
"Semi-accurate" is an appropriate name. He's sometimes right.
Amusingly they are suing Qualcom for doing the same thing to them in the mobile market.
Also, I need to bring this up every time we have an "AMD is destroying Intel" post: Intel's per-core performance is still higher than AMD, somewhere around 10% depending on the application.
AMD is competitive again, which is great. But I'm still buying Intel because my use-cases demand maximum per-core performance.
There are many use cases, and maybe to you per-core performance is paramount, but he's making the case that for the market Xeon-Ws are aimed at, cores are more important than 10% peak performance, otherwise you wouldn't be buying a 36 core monster to begin with.
Yes, in pure per core performance intel still has a nice edge of roughly 10%, and yes games mostly care about single core perf, and yes if you want to make the best gaming machine money can buy you still go for intel.
But gaming has not been CPU bound for quite some time, and amd's proposition there is not "the same thing, little bit less expensive and little bit less powerful", it's "the same price, little bit less powerful by core but still not cpu bound, and we give you twice the cores". Intel might very well fix that by the end of the year but right now you want a 600 - 1000 € gaming machine there is a very good chance you go for Ryzen.
And of course that's why AMD did that, they did not go crazy with core count to be nice, they did it because that's what they finally needed to get over that "but intel's still ahead and there is no reason to go with amd if you don't need to save fifty bucks".
What I like about AMD is that they pack their chips with features. Low end ones support VT-D and VT-X, AES, etc., whereas Intel likes to segment the market a ton with arbitrary features turned off to prevent higher ticket SKUs from being replaced with lower priced ones.
Alas, strategy games while CPU bound are not fully capable of using any number of threads -- Cities: Skylines tops out around 8 threads, for example.
All these latest generation chips boost clock speed ("turbo" mode) the fewer cores there are in use. AMD is far more generous with its turbo modes at multiple tiers (read: higher percentage of top turbo speed at the same active core count) although at a significant TDP cost
1 - https://www.phoronix.com/scan.php?page=article&item=amd-tr-1...