AMD Threadripper 2990WX Linux Benchmarks: The 32-Core / 64-Thread Beast

[chithanh]

Any chance we can get some ffmpeg H.264 and H.265 encode testing? How about 7z or tar+pxz compression testing?

Computerbase did some 7-zip, Handbrake, x265 testing: https://www.computerbase.de/2018-08/...in_anwendungen (in German)

Do note, they tested on Windows and they noticed serious scaling problems which they could not find the reason for. It might very well be that Windows is to blame for the problems rather than the application or the CPU.

[haasn]

It would be cool to know if the new threadrippers still suffer from the memory latency issues of the current ones. The tl;dr is that under heavy memory pressure (e.g. `stress --vm 32 --vm-bytes 256M`), the CPU cores can stall for many-ms intervals, even up to 1000ms (!) even on a low-latency kernel. (As tested by commands such as `cyclictest`). The problem seems to affect all Ryzen CPUs, including the newer Ryzen 2 ones; but it gets worse on the threadrippers by an order of magnitude or two. (Although disabling NUMA mode removes the insane spikes at the cost of everything else being slower)

If this is the still the case on the newer threadrippers I can't seriously recommend them to anybody with a straight face, the CPUs have been riddled and riddled with bugs. (First the segfault issues, then the PCI reset debacle, the insane CPU stalls in NUMA mode, the C6 state freezes on the newer AGESA/UEFI versions, etc.). I wish AMD would just go ahead and fix their broken CPUs already before making more of them.

(cf. https://old.reddit.com/r/Amd/comment...hread/e3aot8z/)

[tchiwam]

Some benchmark site shows some memory bandwidth limitations on the 2990WX, would be interesting to find similar memory bound cases on linux and see if Windows will simply have to upgrade their scheduler do get better performances.

[thelongdivider]

I think Linux is performing substantially better than Windows with these CPUs. The memory bandwidth is concerning based on other reviewers.

What about results in Linux gaming? I don't know of any native Linux games able to use so many cores, but maybe some DX11 titles on Wine + DXVK could benefit of them (GTA5 is on my mind specifically).

Definitely no current game will use 64 threads. Most of them don't even use more than 8 threads on any operating system!

[Michael]

What about results in Linux gaming? I don't know of any native Linux games able to use so many cores, but maybe some DX11 titles on Wine + DXVK could benefit of them (GTA5 is on my mind specifically).

I have some Linux gaming CPU core scaling results coming up tomorrow or so with the GTX 1080 Ti and Vega 64.

[haasn]

It would be interesting to see if they still haven't fixed the memory latency spikes when the machine is under VM pressure. (Orders of magnitude worse than any other machine I've ever seen)

An easy way to test this is to run `cyclictest --numa -p90 -m` at the same time as `stress --vm 32 --vm-bytes 256M`. If you see the maximum latency spiking into the 10,000 range and higher, they still have the same issue. I've even seen spikes as long as 2 seconds, during which the CPU essentially just freezes. (Check any Intel or pre-Ryzen CPU and you will get results in the 1-5ms range on a generic kernel, and in the 100-500μs range on a lowlatency kernel)

The only work-around seems to be either disabling NUMA mode (which brings it down to high, but more acceptable values) or isolating CPUs on which you intend to run realtime/interactive tasks to prevent them from experiencing memory pressure.

A script like that could be built with a few lines of shell, and would make a very interesting graph to include in your results - maybe AMD will then notice this issue (that affects all ryzen systems I've seen) and ignore it instead of sweeping it under the rug as usual.

[Michael]

Some benchmark site shows some memory bandwidth limitations on the 2990WX, would be interesting to find similar memory bound cases on linux and see if Windows will simply have to upgrade their scheduler do get better performances.

I haven't seen any major memory bandwidth shortcomings yet in any of my tests and have more tests on the way.

[yro84]

So, this cpu is a monster.. a ripper... the name says everything

Nice to see AMD back in business and pushing intel to the corner. Big fan of AMD products and very happy with what i have been seing in the past few months..

[bridgman]

It would be interesting to see if they still haven't fixed the memory latency spikes when the machine is under VM pressure. (Orders of magnitude worse than any other machine I've ever seen)

Do you mean "worse than any other NUMA system" or just "worse than a non-NUMA system" ?

[tildearrow]

Typo:

unless you plan on multi-trasking or carrying