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

[csbin]

AMD EPYC Infinity Fabric v. Intel Broadwell-EP QPI Architecture Explained

https://www.servethehome.com/amd-ep...ntel-broadwell-ep-qpi-architecture-explained/

 

[krumme]

16x of cheap 32GB for 512 is gona sell in itself

 

[JoeRambo]

16x of cheap 32GB for 512 is gona sell in itself

Those 512GB at current prices and massive BW is HUGE asset in current memory market.

Lots of advantages for AMD, no doubt about that. 8 channel mem, decent memory speeds, lots of processing power and platform connectivity.

And all that when Intel is releasing Skylake-SP, that was formulated by marketing department. Insane segmentation, lowered memory speeds, lowered link speeds, lowered AVX512 throughput. L3 cache neutered.

The only thing "Scalable" about those CPUs is going to be Intel's market share and bottom line. Both are going down hard.

The only thing keeping my company from buying Epyc stuff for some of our tasks is that stupid GCC bug, AMD really needs to fix it pronto.

 

[ajc9988]

http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade

 

[ub4ty]

AMD EPYC Infinity Fabric v. Intel Broadwell-EP QPI Architecture Explained

https://www.servethehome.com/amd-ep...ntel-broadwell-ep-qpi-architecture-explained/

These guys have had Epyc for as long as anyone else in a functional system and had the ability to run benchmarks and publish results.
Anandtech and others beat them to the punch as well as stacked EPYC vs Intel's latest offering.
Now, it's July 11th and their website is littered with information on the Intel Xeon offering yet is void of Epyc information which they have had up and running in their labs.

This is inexcusable and appears fishy IMO.

 

[moinmoin]

These guys have had Epyc for as long as anyone else in a functional system and had the ability to run benchmarks and publish results.
Anandtech and others beat them to the punch as well as stacked EPYC vs Intel's latest offering.
Now, it's July 11th and their website is littered with information on the Intel Xeon offering yet is void of Epyc information which they have had up and running in their labs.

This is inexcusable and appears fishy IMO.

I like to give them the benefit of doubt and trust them saying that the platform is still in flux with bios updates and so on (happened as well with Ryzen on AM4 after all). But them pushing out that avalanche of Xeon articles while claiming to hold back Epyc results for trying to be fair to AMD is definitely a headscratcher.

 

[ajc9988]

I like to give them the benefit of doubt and trust them saying that the platform is still in flux with bios updates and so on (happened as well with Ryzen on AM4 after all). But them pushing out that avalanche of Xeon articles while claiming to hold back Epyc results for trying to be fair to AMD is definitely a headscratcher.

These guys have had Epyc for as long as anyone else in a functional system and had the ability to run benchmarks and publish results.
Anandtech and others beat them to the punch as well as stacked EPYC vs Intel's latest offering.
Now, it's July 11th and their website is littered with information on the Intel Xeon offering yet is void of Epyc information which they have had up and running in their labs.

This is inexcusable and appears fishy IMO.

Even Anandtech appears to be slammed at Intel in places, but they disclose having the Epyc platform only in one week, that it was in flux, etc., AND STILL showed AMD going toe to toe in many tasks, and even winning in some. It even concluded the equivalent priced chip would be outperformed by AMD, trying to be well rounded in its statements (took issue with some of the ram coverage, like latency and what that meant for performance). But, in a market with 99% Intel server share, some bias is to be expected.

 

[ub4ty]

I like to give them the benefit of doubt and trust them saying that the platform is still in flux with bios updates and so on (happened as well with Ryzen on AM4 after all). But them pushing out that avalanche of Xeon articles while claiming to hold back Epyc results for trying to be fair to AMD is definitely a headscratcher.

Didn't seem to stop Anandtech from getting the Epyc benchmarks executed and posted...
Furthermore, there didn't seem to be an embargo in place restricting anyone from publishing results. Thus, that isn't an excuse either.

They had wall to wall coverage of the intel system + benchmarks + pictures of the chips and inside the case and all sorts of eye candy .. Well over 4-5 major posts. Yet, they mention a blurb at the end about core latency and what is to come when they finally get around to detailing the (2nd class citizen - Or so they've placed it)....

If anything, Intel had an influence on delaying Epyc detailing until after their product launched and was detailed so as to place AMD 2nd.
I don't like it at all and there is no reason for it.

Even Anandtech appears to be slammed at Intel in places, but they disclose having the Epyc platform only in one week, that it was in flux, etc., AND STILL showed AMD going toe to toe in many tasks, and even winning in some. It even concluded the equivalent priced chip would be outperformed by AMD, trying to be well rounded in its statements (took issue with some of the ram coverage, like latency and what that meant for performance). But, in a market with 99% Intel server share, some bias is to be expected.

Anandtech got the AMD benches out. Even with it being 'unstable', it bested the intel platform in several areas.
Were all grown adults.. In this current environment, we can handle : the platform is still being stabilized with BIOS/firmware updates and performance will likely improve.

Suffice to say, I now have the information I need to further evaluate what choices I will be making w.r.t to purchasing. This served as good detailing for Epyc and a considerable amount can be surveyed about Threadripper from here. Intel's HEDT/serve offerings are now detailed as well. Lots of data and information to go over and some guesstimates to be mad.

Thank you very much Anandtech

 

[StefanR5R]

If anything, Intel had an influence on delaying Epyc detailing until after their product launched and was detailed so as to place AMD 2nd.
I don't like it at all and there is no reason for it.

No need to assume that Intel tried and succeeded to pressure these sites to defer EPYC coverage. Skylake-SP has been in use in datacenters at certain large customers for several months already (I suppose production use). EPYC on the other hand only got fresh out the gate just now.

 

[formulav8]

Insane segmentation

That's an Intel registered trademark.

 

[ajc9988]

No need to assume that Intel tried and succeeded to pressure these sites to defer EPYC coverage. Skylake-SP has been in use in datacenters at certain large customers for several months already (I suppose production use). EPYC on the other hand only got fresh out the gate just now.

False. Many large companies that are AMD partners, like M$ Azure, are partners, have used this since March or so (compared to June for Intel), and have done ad videos saying as much. So, please, don't lie.

 

[Markfw]

False. Many large companies that are AMD partners, like M$ Azure, are partners, have used this since March or so (compared to June for Intel), and have done ad videos saying as much. So, please, don't lie.

And your proof is ? Link ?

 

[ajc9988]

And your proof is ? Link ?

https://www.geekwire.com/2017/microsoft-azure-baidu-embrace-amds-new-epyc-data-center-processor/

I'll find more links, but that is one of the first when you search Microsoft azure Epyc. You then have the promotional videos from AMD on release date talking about them evaluating it for months. Do you want links for every part of my sentence?

 

[ajc9988]

https://www.theregister.co.uk/AMP/2...as_as_microsoft_armlessly_embraces_new_xeons/

List of people sampling pre-release Epyc:
https://venturebeat.com/2017/06/20/amd-launches-broad-epyc-server-line-with-up-to-32-cores-per-chip/
https://www.bloomberg.com/news/arti...-chip-revival-effort-enlists-some-big-friends (includes list of uses at Baidu)

 

[ajc9988]

And your proof is ? Link ?

https://youtu.be/Qzz5EsAMdBY
https://youtu.be/lIXTm63e974
https://youtu.be/Fq9gRQWzE1o
https://youtu.be/xf3U9ig7cj4
https://youtu.be/uThuZtttjpY
https://youtu.be/D8IQjjsghEI

I don't want to be accused of posting advertising, I was asked for proof of my statement. These testimonials talk about using it for months, with these videos made for the June 20th event. Intel started demoing around June 1 according to the register article.


Edit: one last article on Dropbox evaluating released at the May 16th investor day:
https://venturebeat.com/2017/05/16/amd-sets-june-launch-for-zen-based-naples-server-chips/

 

[StefanR5R]

No need to assume that Intel tried and succeeded to pressure these sites to defer EPYC coverage. Skylake-SP has been in use in datacenters at certain large customers for several months already (I suppose production use). EPYC on the other hand only got fresh out the gate just now.

False. Many large companies that are AMD partners, like M$ Azure, are partners, have used this since March or so (compared to June for Intel), and have done ad videos saying as much. So, please, don't lie.

Yes, partners have already been working for a while now with AMD's engineering samples to develop boards and servers. But where do you take "June for Intel" for this stage from? June, or end of June, was when Intel sent final hardware to press for tests with embargo till June 11.

Remember that Intel showed a Skylake-SP cluster based on engineering samples at SC16 in November 2016.
https://www.heise.de/newsticker/mel...chste-Xeon-Generation-Skylake-EP-3465742.html
Also in November 2016, Intel began shipping preliminary Skylake-SP to partners.
https://www.servethehome.com/live-intel-ai-day-2016/
In February 2017, Google opened access to Skylake-SP in their cloud.
https://www.servethehome.com/skylake-ep-now-available-though-google-cloud-platform-program/

 

[moinmoin]

Going back to ServeTheHome's fishy reluctance to release benchmarks on Epyc, Ryan Smith just confirmed that their Epyc server was tested as-configured by AMD. Meanwhile Patrick Kennedy from STH gives insight of the progress on benchmarking Epyc.

I think this may be final firmware. Still doing some testing.

Looks like the box we have is getting an upgrade later this week. The good news is that will push us to DDR4-2666 which is important for IF as well.

Hopefully the new configuration we can publish benchmarks on. I have two sets of charts for the Skylake launch. One with 2x EPYC 7601 + DDR4 2400 and one without.

AMD is sensitive to us publishing numbers on de-tuned platforms, which I understand and agree with given it is a new platform and we are still a few weeks from folks being able to buy them publicly.

It is one of the hard bits about the work and a lesson learned from Ubuntu 14.04 with ThunderX that we are carrying forward. I do not think it helps anyone publishing pre-release system benchmarks if there is an expectation that those numbers can change.

Also, STH is not using an AMD reference platform tuned by AMD engineers so we are going more for real-world performance.

I'd also like more numbers now. But if this results in more thorough benchmarks it should be worth the wait. It's already clear Epyc is competitive with Skylake-SP, so better details in what real life areas it excels and where's room for improvements are always appreciated.

Meanwhile the Skylake-SP launch is in danger of being overshadowed by the surprise reveal of Ryzen 3 and 7900X-beating Threadrippers.

 

[scannall]

I'm not sure if anyone has seen this yet. But the project 47 Epyc Petaflop demo looks pretty incredible. Project 47

 

[IEC]

I'm not sure if anyone has seen this yet. But the project 47 Epyc Petaflop demo looks pretty incredible. Project 47

1 giant room full of server racks in 2008 = 1 petaflop

Today, with Epyc + Vega Instinct they can fit that horsepower in a tiny fraction of the space. Amazing indeed.

 

[tamz_msc]

Heise.de publised their SPEC2006 results for Platinum 8180(28C/56T), Epyc 7601(32C/64T), Gold 6148(20C/40T) and E5-2699 v4(22C/44T), all of them in 2P configuration.
https://www.heise.de/ct/ausgabe/201...6148-E5-2699v4-und-AMD-Epyc-7601-3787494.html
All measurements with 64-bit code, without auto-parallelization, without SmartHeap with Intel System Studio 2017Up4 (opt: AVX512 / AVX2)Or x86-open64 4.5.2.1 (AMD, opt for bulldozer / piledriver).

It's paywalled(1.5 Euro). I've had a look at it and the most important findings are:

• 2X Epyc 7601 and 2X Xeon Platinum 8180 are extremely close. Xeon was 12% ahead in int_rate_base and 4.5% ahead in fp_rate_base.
  
  
• Without NUMA optimiztion, Epyc is ahead in STREAM-Triad for very low thread counts(<8) sometimes by almost 2X over Skylake-SP Xeons. The 8180 is bandwidth-starved compared to the 6148 by 35% for 16-32 threads. When fully saturated with the maximum permissible number of threads, bandwidth on Skylake-SP dips while Epyc remains constant.
  
  
• The huge performance disparity in libquantum is obliterated when compiled for 64-bit. Compile times were reduced to two seconds from 10 minutes on 32-bit mode. Yes, minutes. Epyc and Xeon 8180 give the same performance even with AVX512 enabled on the latter in 64-bit mode.

 

[tamz_msc]

SPEC2017 results by heise.de are out:

• Something is wrong with the integer performance on EPYC on the 'speed' benchmark(which measures time to completion for a single job) - 2x Xeon Platinum 8180 is over 60% faster than 2x Epyc 7601 using intel17 -O3. Same issue with gcc, so might be a compiler issue. AMD says their own AOCC does fine, but not tested due to time constraints.
  
  
• This discrepancy is not there when running the 'rate' benchmark(measuring throughput running multiple copies of each job) - Xeon Platinum is only ~5% ahead of Epyc with intel17 -O3.
  
  
• There is no issue with fp performance however. In the 'speed' benchmark Xeon Platinum is just ~4% ahead of Epyc. However, in the throughput-oriented 'rate' benchmark Epyc pulls ahead of the Xeon Platinum by ~16%, again with intel17 -O3. Due to the huge memory bandwidth it offers, in some of the benchmarks, like fluid-dynamics using the lattice-Boltzmann method, Epyc is over 55% faster than Xeon! Epyc does very well in general with throughput-oriented numerical PDE solvers from what I've seen, which is expected because these kinds of HPC workloads absolutely love memory bandwidth and generally run without issues when made NUMA-aware.
  
  
• AVX512 does nothing over AVX2 with the Intel compiler; AVX2 is still slightly faster for the 8180, however with the 20C Gold 6148, AVX512 is marginally faster than AVX2 in the 'speed' benchmarks, which suggests that there is some kind of bottleneck with the higher core count chip; which is ironic given Intel's claims with the mesh topology and cache restructuring, but not entirely unexpected. On gcc 7.1 however, skylake-avx512 opt. flag is behind broadwell opt. flag with the 8180, sometimes by a long shot with the rate benchmark.

Bottom line(my interpretation): it's down to application preferences when it comes to which processor you're going to buy. There are still issues with int performance on Epyc, so Skylake-SP is the winner integer workloads as of now. However, Epyc is obviously much faster in fp performance, and since AVX512 does nothing at the moment for Intel, it is obviously the better choice.

 

[krumme]

I think its fairly easy to do tco calculations for those different solutions.
They are both brand new solutions and obviously there is still some work to do but its probably fixed before they get sold and used so its down to what we know. Eg. Wide vectors or mem bandwith, bandwith/cost. If your loads is heavily sided here you dont excactly need more than a napkin. Goes for much of it as the general efficiency seems so similar its easy to do just by looking at perf/cost for cpu-memory and then check on space/cooling. I think this rough estimate hit spot on for 95%. So if there is a solid difference there is no need to do an expensive comparison on investment alternatives. Saves cost and time at a miniscule risk.

 

[Ajay]

Bottom line(my interpretation): it's down to application preferences when it comes to which processor you're going to buy. There are still issues with int performance on Epyc, so Skylake-SP is the winner integer workloads as of now. However, Epyc is obviously much faster in fp performance, and since AVX512 does nothing at the moment for Intel, it is obviously the better choice.

If you are going to run AVX-512 workload, you'll probably use an optimized library that does make proper use of vector computations.

 

[tamz_msc]

If you are going to run AVX-512 workload, you'll probably use an optimized library that does make proper use of vector computations.

What more optimization do you need than -O3 and -avx512 on the Intel compiler? It doesn't help at all on real workloads, at least not in its present state. In many cases you're going to be better off with explicit vectorization than relying on compiler autovectorization, which is basically the whole point of avx.

 

[TheGiant]

So what is the bottleneck of SKL-SP with AVX512? Memory bandwitch? L2 cache, mesh frequency? L3 cache?

I think it could be tested with 18C SKL-X and DDR4 2666 as a baseline what is really behind it as you can change DDR4 frequency, mesh frequency and other settings you cannot with xeon

I don't believe avx512 is utterly useless....