AMD EPYC Server Processor Thread - EPYC 7000 series specs and performance leaked

[formulav8]

I still do not see the reason why AMD did not just use their own optimizing compiler

Probably because no one really uses it unlike Intel's.

 

[lolfail9001]

For the same reason that you or I did not bench these CPU's with our own optimizing compiler; we don't have the resources to compete with Intel. AMD is competing in the CPU market here and carefully allocating its limited engineering resources.

Here's what you forget: AMD did optimized SPEC runs anyways, and they beat Intel with a solid margin anyways. Hell, they didn't even use their patched LLVM for it, they used their own compiler that only supported bulldozer for it. I am not kidding, look at it http://spec.org/cpu2006/results/res2017q2/cpu2006-20170529-47126.html .
AMD used these resources anyways, so your paragraph does not even read like an excuse, simply because it does not reflect reality.

Probably because no one really uses it.

Much better argument, but nobody uses excel for performance testing of CPUs either, or does one?

 

[rainy]

Quote where they said they are buying them please.

You still don't understand marketing speak.

Anyway, of course they may buy some Epyc Servers. The question is how many.

Do you think the major cloud providers will making a move away from Intel? In what timeframe?

I must openly admit, that observing your butthurt is a pretty epic thing./s

 

[Ajay]

@Markfw I think the initial game afoot here is just smart business. Large cloud providers and data centers need to buy Epyc based systems in order to pressure Intel so they can gets system costs down. What ever risk there are in moving to AMD's new platform, they are clearly worth it. I think this suggests that the risks are not too high and that the performance is sufficient for now (I would guess these large companies have/are doing extensive testing on their application stacks).

I think this is how the stock analysts have arrived at their initial conclusions. They are more negative on Intel (margin decline) than they are positive on AMD (unit sales). The $10,000 question (well, billion dollar really) is how sales and profits are for AMD and its vendors are in a year, two years, three, etc.

Personally, without knowing actual comprehensive data on performance and efficiency for various application stacks, I think AMD has a solid product to build on. Near term execution on ISV/IHV support, integration and expanded marketing will play heavily on AMD's performance in the server space that they target. Interesting times for sure.

 

[Topweasel]

@Ajay. The Dell support is legit. Now that they are Private considering what Intel did last time they tried to offer AMD products. I don't think Dell's support of Ryzen and ThreadRipper is to force Intel into better pricing. They don't have to worry about the bottom line as much. I see their support while being worthwhile a little bit more long haulish as the would want to prop up an actual competitor against Intel to avoid being strong ARMed again.

Also with Azure. I don't think MS cares enough to take an actual hit in performance scalability to adopt something that wasn't competitive in the market. But I see them maintaining Ryzen purchases even if Intel does feel pressured to lower prices because AMD has been a great hardware partner to them for a while and giving AMD some stability here where the products are life and death to AMD, keeps AMD alive to support their other hardware initiatives (Xbox).

The real question for the next couple years though will really come down to supply. AMD can't not sell enough this year to really impact Intels shipments. But if they can get their suppliers to ramp up. Next year and on becomes the real battle in terms of AMD's legitimacy.

 

[Markfw]

Now there one other thing here. I am only guessing based on known facts, but the Ryzen/TR/EPYC systems seem to have a efficiency lead. and saving (just a random number to reference) $1000 a month in electricity on the box may also save $1000 on AC, and even if they are 10% slower than Intels similar CPU, between the product cost (most likely much lower) and the energy savings could also be a big factor.

Only time will tell (as stated above)

 

[krumme]

Uhm, since when is 2/3rds of the die are CCXs? Die size is ~195mm^2, 2 CCXs total 88mm^2. Unless i've been missing on something, 195*2/3=130, if anything 2/3rds of the die are L3+non-CCX parts. Similar with split.

Of course in the end it does not quite matter in this math because it is a difference of 4 dies, or about 1.3% if i assume that otherwise it was flawless and sure at such defect density 95-97% yield sounds reasonable. But difference between 97% and 99.9% yield is actually massive and i hope we all understand why.

P. S. where the hell did looncraz pull that defect density figure from?

Did you click Fots 99.9%?
I didnt.
25 or 40usd a die doesnt matter squad in a 2000 usd cpu so i couldnt care less.
The point still stands. And i wouldnt be surprised if its +97% looking at the portfolio as you are. So who cares about that tweet.

I dont know how real efficiency pans out but seen from here avx512 huge dies with tons of ressources running idle vs 189mm2 dies seems like a bad business idea.

Its working optimal when you have monopoly to cover all segments and positions with a monolitic solution but when someone suddenly arrives at the party with a lean product that seemingly addresses say 60% of the market more effectively and have apple and samsung as future and current process sponsor - then there is bound to be trouble.

It doesnt take a genius to see that and thats probably why the investors react as they need it very obvious to react when its not solid financial numbers or forecast.

Zen was performing much much faster in desktop that anticipated. Inter die and socket bandwith is plenty. Inter ccx latency is fine and comparable to Intel mesh in a core. Damn impressive btw.

We have yet to see anything that doesnt point to a success.
I guess share 2018 7% 2019 11% 2020 17%

 

[IEC]

First motherboards being listed:
https://www.supermicro.nl/products/nfo/AMD_SP3.cfm?pg=MOBO

The flagship supports the following:

• Supports up to 4TB Registered ECC DDR4 2666MHz SDRAM in  32 DIMMs

That's a lot of RAM...

 

[Ajay]

Now there one other thing here. I am only guessing based on known facts, but the Ryzen/TR/EPYC systems seem to have a efficiency lead. and saving (just a random number to reference) $1000 a month in electricity on the box may also save $1000 on AC, and even if they are 10% slower than Intels similar CPU, between the product cost (most likely much lower) and the energy savings could also be a big factor.

Only time will tell (as stated above)

Exactly! But this is ATCAO, so we will debate the unknowns for months

 

[Jan Olšan]

I don't really get it. It seems some people are really mad that AMD has produced a compelling product.

That happens a lot... What I find most amusing is when some people actually bash the product of the "opposite company" in proportion to the quality/attractiveness.

When the enemy's CPU (for example) is comparatively bad, they are happy, content and don't feel the need to take up arms. However when it turns out better, suddenly they feel challenged or endangered and their activity shoots up.
I saw/see this with Ryzen a lot, people desperately finding things to criticise (and by that I don't mean rational criticism of actual weaknesses, which is of course fine and needed), cherry picking benchmarks or metrics that look bad, investing time to find problems... lots of those people I would not see in comments after AMD-related articles before, they only started to care after Ryzen. I noticed a case or two where a strong fan of Nvidia GPUs started to regularly rant/bash under Ryzen articles probably because it's made by Nvidia's competitor. I assume he was afraid AMD would have more money to compete against his precious

Being fan and rooting for something is not a particularly wrong thing, but when you start to act like that, it means you got too absorbed and need to relax.

 

[csbin]

https://www.servethehome.com/amd-epyc-7000-series-architecture-overview-non-ce-ee-majors/


EPYC is B2 stepping

 

[lolfail9001]

Oh, i have just noticed why exactly STH did not publish their review. Apparently in spite of launch it still goes through BIOS polishing. Uhm, AMD?

Trolling is not allowed
Markfw
Anandtech Moderator

 

[WhoBeDaPlaya]

First motherboards being listed:
https://www.supermicro.nl/products/nfo/AMD_SP3.cfm?pg=MOBO

The flagship supports the following:

• Supports up to 4TB Registered ECC DDR4 2666MHz SDRAM in  32 DIMMs

That's a lot of RAM...

Imagine a Beowulf cluster of those!
I'm sure that would run Quake / Crysis

 

[Kenmitch]

Oh, i have just noticed why exactly STH did not publish their review. Apparently in spite of launch it still goes through BIOS polishing. Uhm, AMD?

Their articles end with the following....

We will have more information on AMD EPYC as soon as we can release it

We will have more information on AMD EPYC including benchmarks once we are allowed to release the information.

Bios polishing? Didn't see it mentioned.

 

[stockolicious]

I don't think Dell's support of Ryzen and ThreadRipper is to force Intel into better pricing.

this is the million dollar question - is this the same ol same ol with AMD becoming competitive - its not - a server vendor now has to ask - should i favor intel when i have AMD as a partner who can supply me a great server product and high end GPU tech - the ability to have a high end CPU and GPU can cause a much higher attach rate not only for Ryzen and vega in gaming rigs but also in the cloud with their new products to be announced on the GPU side next week. AMD has a pricing advantage on both ends a server with all AMD product will be half the price of an INTC?NVDA option. there is a pricing advantage here because Intel and nvidia really sell at a premium and at the other end of the spectrum AMD with infinity fabric leverage great yields and can sell them way cheaper with good margins.

 

[stockolicious]

first - sorry as this seems to be a reply but it will be a different thought - I heard someone on a forum say that the real problem for INTC is if the demand for the 32 core EPYC is high. Intel only plans i think a 28C chip with some other issues. The poster was basically saying if the market wanted 32 core chips or even 28c chips Intel would have supply problems due to their approach in making them. If that is the case this is a design issue that they cant overcome without a redesign - keep me honest here as im more of a business model guy rather than someone who has much tech knowlege.

 

[formulav8]

Much better argument, but nobody uses excel for performance testing of CPUs either, or does one?

me no know

 

[Kenmitch]

Will be interesting once we've seen all things Zen.

 

[Markfw]

Regardless of the results, I think by now we all know "the race is on" and competition is back in the CPU world, personal all the way up to server. (Ryzen to TR to EPYC)

And the only winners are us, the customers, including businesses.

 

[Ajay]

Curious what B2 will mean for Zen. Obviously fixes for Server Ops, but anything else?

 

[formulav8]

When the enemy's CPU (for example) is comparatively bad, they are happy, content and don't feel the need to take up arms. However when it turns out better, suddenly they feel challenged or endangered and their activity shoots up.

There's a member here who even started a Ryzen thread last year based on SR numbers of a very early eng sample that didn't look very good bench wise.

It was updated with Dec results only afterwards but he wouldn't update the thread anymore after it was found out Ryzen wasn't bad after all. Even when persons asked him to update he wouldn't. So someone else ended up starting another thread for SR.

But the whole time and even still, he's very active with his nearly year and a half long Intel threads. Its really weird.

 

[formulav8]

And the only winners are us, the customers, including businesses.

Yes sir

 

[knutinh]

Throwaway account...
I'm a software engineer working on the code optimizer in one of the major C++ compilers out there and I think what AMD did with the results is justified. The Intel compiler is indeed "optimized for SPEC" - it employs optimizations that are either illegal in a language such as C++ or not applicable to pretty much any real program outside the SPEC benchmarks that is larger than a few hundred lines of code. You have to use a magical combination of flags to get anywhere close to the numbers they publish, and if you try that on other programs you either don't see any improvements or might introduce runtime bugs. GCC is overall the most suitable compiler for systems and server software right now.

Every compiler is in a way or another optimized for certain benchmarks because that's usually the code used to test how good the optimizations are, but Intel does seem to go a bit too many steps in this direction. In a way I can understand that they want to make their platform look better by any means...

Comparing hardware when the platform compiler is different (think Mac, Linux, Windows) is also an issue for other benchmarks, such as Geekbench - just from compiler optimizations you can get 10%-20% improvements on the same hardware in many of the sub-benchmarks.

My experience with icc and gcc suggests that I would use icc if compiling numerically heavy code to run on Intel hardware that I had to pay for.

The alternative with gcc might be to pepper the code with inline assembly in order to get vectorization. Bug-prone, non future-proof and resource demanding.

For comparing pure hw performance, one would expect AMD and Intel to have the inhouse resources to optimize a given simple piece of code using whatever means so that it passes a set of tests (not VW style) while being close to optimal fr their hw.

k

 

[Rngwn]

If it was, it was not Intel's complaint/excuse.

I personally am still keen on opinion that Ryzen is a hypervisor CPU first and foremost.

EDIT: Oh, i missed it, but there were some binary benchmarks run as well. https://www.heise.de/newsticker/meldung/AMD-Epyc-legt-los-3748615.html


So, we can even compare to how it stacks up to Skylake-SP now.

EDIT: Fun stuff, it manages to lose a couple of 24 core Skylakes in cinebench.

Here is the "source" of the Platinum 8180 score.

http://www.coolaler.com/threads/xeon-platinum-8180-56112.345288/

Spoiler: chinese text

看到中國的網友對於上次Platinum 8180的跑分表示疑問，會不會是溫度造成降頻

今天把系統重新安裝上去再跑一次，用原本的測試環境溫度直接壓到破表降頻

因為機架沒推到機房內吹冷氣，噴出來的根本是焚風，會燙手..

這次205W TDP真的不是蓋的，絕對足瓦...

換上四顆12CM 8000RPM 才把溫度壓在80度以下，讓時脈維持在3.2GHz

R15分數為「8301cb」，系統上面還有一些服務，可能會讓分數下降，但幅度不會太大

分數可以跟 E7 8890v4 4way (96C、192T)一拚了 2.1GHz 8736 cb、1.9GHz 8299cb

Which can be roughly translated, with my rusty knowledge in Chinese to:

Spoiler: translated text

I just found the most recent, yet questionable Xeon Platinum 8180 benchmark from a Chinese user. The benchmark score itself might be plagued by the thermal throttling.

Today, I re-installed and re-bench the system with the same ambient temperature that is supposedly caused the thermal throttling. The room used for testing is not air-conditioned. Therefore the air used to cool the rack will be quite toasty. Now, the 205W is not all that great, it's simply too much. You need to put four 12cm fans spinning at 8000RPM to keep the temperature below 80c and to maintain the all-core boost at 3.2GHz.

The CB15 score is 8301cb. There might be a bunch of services running so it might interfere with the benchmark. However, do not expect that the score will change much from these background services. This benchmark score is in line with 4x E7 8890v4 (96c/192t) which scored 8736cb while running at 2.1Ghz and 8299cb while running at 1.9Ghz.

Judging by the poster's tone, he seems to be disappointed with the performance and thermals he get with the Skylake-SP. So, it could be that the Skylake-SP in question might also be quad-socketed rather than dual.

So the comparison might very well be 2x Epyc vs 4x Platinum (with badly throttled performance), but don't quote me on this.

 

[DDH]

There's a member here who even started a Ryzen thread last year based on SR numbers of a very early eng sample that didn't look very good bench wise.

It was updated with Dec results only afterwards but he wouldn't update the thread anymore after it was found out Ryzen wasn't bad after all. Even when persons asked him to update he wouldn't. So someone else ended up starting another thread for SR.

But the whole time and even still, he's very active with his nearly year and a half long Intel threads. Its really weird.

Ha, I had to go back and have a look. I guess who it was before I found the thread. No surprises. Having a read through that thread is hilarious the amount of people that took those numbers
and extrapolated AMDs death. And not updating the thread when new information became available... .well thats just like sweeping it all under a rug