Initial Benchmarks Of The AMD EPYC 7601 On Ubuntu Linux

[bsp2020]

Interesting review. Could you also turn on the NUMA interleaving policy for the Xeon system? It only requires 2 NUMA nodes due to Intel's superior architecture but I'd be interested to see if it has any effect.

I'd like to see this as well. Does it make no difference in Intel setup?

[torsionbar28]

Interesting review. Could you also turn on the NUMA interleaving policy for the Xeon system? It only requires 2 NUMA nodes due to Intel's superior architecture but I'd be interested to see if it has any effect.

Ok, but your "superior architecture" got trounced in most of these benchmarks. Any modern server or hypervisor OS running on this beast is going to be NUMA aware.

[Apopas]

Michael, I demand screenshots from the boot process with all the many penguin logos, power numbers and if possible timed compilation of LibreOffice. I want to see what I am missing...

I had never thought about that! It indeed displays 80 penguins for Xeon???

[cb88]

Michael are you sure the AES benches are correct... even a Ryzen 7 trounces all of it's competitor chips by about 30% when running the AIDA64 AES benchmark... and AMD does support AES-NI since bulldozer I think. It's probably worth looking into to figure out what the deal is there...

[Mark Rose]

Yowsa this AMD Epyc is an intel killer! Dual socket Xeon Gold can't keep up with a single Epyc in most tests. Imagine what a dual socket Epyc can do... At this rate, we'll surely be seeing Epyc appear on the top500 supercomputer list in the coming years. AMD share price $25 this time next year?

Not necessarily. Intel is still king when it comes to floating point throughput. A Ryzen 1800X barely competes with an i5-7600 for what I do. Epyc may do well in some special use cases, like data capture at CERN, but it's not the chip I'd run to for doing weather modelling, nuclear simulation, etc. Mind you, much of that is being done on GPUs, and the extra IO could make Epyc appealing as a host system. Epyc is much more a general purpose CPU than an HPC CPU.

[Spooktra]

Thanks for the article Michael. While I am not as enthused by Epyc's showing as most people in this thread, it's still an interesting read. Any chance of adding a 4k x265 encoding test to these results, this should scale better than the x264 test and is a more intensive encode. Thanks.

[GruenSein]

Not necessarily. Intel is still king when it comes to floating point throughput. A Ryzen 1800X barely competes with an i5-7600 for what I do. Epyc may do well in some special use cases, like data capture at CERN, but it's not the chip I'd run to for doing weather modelling, nuclear simulation, etc. Mind you, much of that is being done on GPUs, and the extra IO could make Epyc appealing as a host system. Epyc is much more a general purpose CPU than an HPC CPU.

With respect to AVX, that is true. In theory, a modern i7 or Xeon has a much wider vector units. In practice, this is rarely relevant since most software is compiled for basic x86_64 arches which require only SSE3 if I am correct. Additionally, software needs to be tweaked a lot to make good use of those units. Setting the -march flag usually doesn't cut it. I work at a university and my department does a lot of CFD (which is pretty comparable to weather forecasting, I assume). Even here, the code is rarely optimized for AVX and such as it takes up a lot of time and the code needs to run on many kinds of HPC machines (some using Power CPUs, some older Xeons). Also, execution units are not the bottleneck in many cases. The code gets starved by lacking memory bandwidth.

If you have the time, a problem that does not require too much memory bandwith, and a very specific deployment target in mind using those SIMD units on Intel does pay off big time, though. However, I would argue that the "special cases" Epyc performs well in are closer to the norm than the ones making good use of the extra SIMD units in modern Xeons.

[pattakosn]

Thanks for the article, as always! It is mentioned that the intel gold cpus cost 1k more than the amd cpu. However I would like to knows the total cost of each server as configured for these tests. eg the memory alone required for the intel system could make an extra difference

[leonmaxx]

basic x86_64 arches which require only SSE3 if I am correct.

SSE2. First x64 CPUs Athlon64 have only SSE2 support.

[foppe]

in reply to an earlier question: note that since there's no SB, lanes are also eaten depending on whether you add one/more NICs to the *board*, USB, (audio), etc. Which is why in practice you'll have <128 lanes to play with. Still, quite a bit of flexibility there, and more than enough lanes to get by.