AMD EPYC On Ubuntu 18.10 Putting Up A Stronger Fight Against Xeon Gold

[ypnos]

I agree that the comparison is interesting but doesn't reflect the decision a buyer would make today. We bought a dual socket system and even then we went for 2x7551 instead of 2x7601. Our system builder of choice charges €3.371,80 per 7551, €4.255,00 per 7601 (listed prices w/o VAT). That's a price increase of 26% for getting the flagship model, which is at most 10% faster.

To put in perspective what others wrote about the single socket system: For a 7551P they charge €2.116,00, so it is half price of the 7601.

It is fair to test what you have at hands, but when doing performance/cost comparisons this should be mentioned.

[Spooktra]

Sorry, but I just don't trust PTS anymore, look at the x265 encoding tests, a dual Xeon Gold 6138's and an EPYC 7601 manage 10-11 fps and the Ryzen 1600 I talked about in other thread manages about 6 fps?!? Before you lose any credibility you may have, I strongly suggest you stop using PTS until you dive deep into what is wrong with it, Xeon Gold 6138 processors are 20C/40T beasts with 6 channel ram and 2 FMA AVX-512 units per core and x265 is known to love memory bandwidth and makes good use of AVX-512, in benchmarks I have seen on other sites, dual port AVX-512 units make Xeons nearly untouchable in tests that make use of that instruction set.

Something is very screwed up with your software.

[chithanh]

It's all a matter of the hardware I have available... No AMD 2P server. AMD talked of sending out a 2P system, but so far haven't seen it.

Not reviewing hardware that you don't have available is of course understandable, I am not criticizing that.

Mentioning the situation would still be the correct thing to do as the other post wrote. Heck, a note that it is AMD's fault that they look so bad due to not sending hardware for review might even gain attention of the folks who make the decisions on who receives hardware for review.

[edwaleni]

Sorry, but I just don't trust PTS anymore, look at the x265 encoding tests, a dual Xeon Gold 6138's and an EPYC 7601 manage 10-11 fps and the Ryzen 1600 I talked about in other thread manages about 6 fps?!? Before you lose any credibility you may have, I strongly suggest you stop using PTS until you dive deep into what is wrong with it, Xeon Gold 6138 processors are 20C/40T beasts with 6 channel ram and 2 FMA AVX-512 units per core and x265 is known to love memory bandwidth and makes good use of AVX-512, in benchmarks I have seen on other sites, dual port AVX-512 units make Xeons nearly untouchable in tests that make use of that instruction set.

Something is very screwed up with your software.

Have you run the x265 test yourself? Michael updates his test files on some releases to match up with a typical use case.

Like when I thought I found a problem in his test of encoding/decoding on Clear Linux, until I broke down his script and found with the new release he came out with a more complex video file so the results didn't match up with prior tests.

PTS is open source, if you think its foobar, take a look at it, I know I do. If you are an advocate for FOSS and you think something is wrong, aren't you compelled to look further?

On those recent Threadripper WX tests with PTS, there were a lot of things that didn't look right, especially in Windows. I went through his test scripts to see what gives.

You are aware of the AVX penalty in Intel CPU's? That AVX performance is not the same on all Xeon's? (Hence the silly branding on Gold, Silver, etc.)

Food for thought.

[Spooktra]

Have you run the x265 test yourself? Michael updates his test files on some releases to match up with a typical use case.

Like when I thought I found a problem in his test of encoding/decoding on Clear Linux, until I broke down his script and found with the new release he came out with a more complex video file so the results didn't match up with prior tests.

PTS is open source, if you think its foobar, take a look at it, I know I do. If you are an advocate for FOSS and you think something is wrong, aren't you compelled to look further?

On those recent Threadripper WX tests with PTS, there were a lot of things that didn't look right, especially in Windows. I went through his test scripts to see what gives.

You are aware of the AVX penalty in Intel CPU's? That AVX performance is not the same on all Xeon's? (Hence the silly branding on Gold, Silver, etc.)

Food for thought.

Yes, I installed and ran PTS on my system as I explained in https://www.phoronix.com/forums/foru...ny-tests/page2.

[Filiprino]

I have in my hands a 4P Opteron 6276 system in which I am running the tests you have run with the 2990WX vs Epyc 7551. They'll be uploaded to openbenchmarking.
Seeing the terminal output, performance shows a clear advantage for the single socket Epyc but there are a number of tests in which the 4P Opteron is better or equal.

I downloaded the Debian package an installed it in an Ubuntu 18.04 LTS. Storage runs from an SSD RAID.

[nuetzel]

Timed Linux Kernel Compilation
Didn't we saw ~30-31 secs with Epyc?

[Michael]

Timed Linux Kernel Compilation
Didn't we saw ~30-31 secs with Epyc?

Depending upon the compiler used and storage device (Optane 900p), yes.