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

[schmidtbag]

Typo:

You're late!

[Mark Rose]

Were these tests done in NUMA or UMA mode?

[tildearrow]

You're late!

I was sleeping while he posted the articles.

[Mark Rose]

Were these tests done in NUMA or UMA mode?

I guess I can answer my own question: the WX chips only run in NUMA mode.

[angrypie]

Incredible performance on Linux, as expected. I'm reading reviews for Windows software and a great deal of them scale like shit.

[schmidtbag]

I guess I can answer my own question: the WX chips only run in NUMA mode.

Sort of - through the Ryzen Master software, you can enable "Game Mode" which is kind of like UMA. I don't think it's exactly the same thing but it's as close as you can get.

[nils_]

Sort of - through the Ryzen Master software, you can enable "Game Mode" which is kind of like UMA. I don't think it's exactly the same thing but it's as close as you can get.

How does that work exactly? Pretending to be a single NUMA node or are they disabling a few nodes? The memory bandwidth / latency issues (as well as PCIe) may be a lot worse with this CPU since it seems to only offer 4 memory channels, meaning either one channel per NUMA node or 2 CPUs not having any direct memory access at all (going over infinity fabric).

It would be interesting to see benchmarks with larger working sets and maybe even some disk I/O.

[nils_]

Unless Intel comes up with something really crazy, and I don't mean anything like their oh-sh*t-we-need-to-cobble-together-something-really-cool-ASAP demo powered by a 1000W water cooler at Computex this year, I think they've basically lost the business from us for the next couple of years.

They could do something like AMD does and just sandwich together multiple CPUs in an MCP design. I don't think they're too worried at the moment.

[schmidtbag]

How does that work exactly? Pretending to be a single NUMA node or are they disabling a few nodes? The memory bandwidth / latency issues (as well as PCIe) may be a lot worse with this CPU since it seems to only offer 4 memory channels, meaning either one channel per NUMA node or 2 CPUs not having any direct memory access at all (going over infinity fabric).

To my understanding, it just reduces the available usable cores down to 1/2 or 1/4, depending on your selection. I think it's safe to assume that the disabled cores are the ones that don't have direct memory access. When running 1/4, it basically turns into a 2700 (dual-channel, 8 cores). I think the reason they did this is because even though the 1/2 selection is still quad-channel, it's 2 channels per die, so you could still end up losing performance by talking across the InfinityFabric.

So - if you want the best possible performance for single-threaded tasks, the 1/4 selection is probably your best bet.

[nils_]

To my understanding, it just reduces the available usable cores down to 1/2 or 1/4, depending on your selection. I think it's safe to assume that the disabled cores are the ones that don't have direct memory access.

Interesting, thanks for clarifying. This seems to mostly deal with limitations in Windows, after all one could just pin the game/application to a single NUMA node.