NVIDIA GH200 72 Core Grace CPU Performance vs. AMD Ryzen Threadripper Workstations

[phoronix]

Phoronix: NVIDIA GH200 72 Core Grace CPU Performance vs. AMD Ryzen Threadripper Workstations

Earlier this month I posted some initial CPU benchmarks of the NVIDIA GH200 Grace Hopper up against AMD EPYC Zen 4 and Intel Xeon Emerald Rapids processors. That was a very interesting battle and showed the interesting capabilities of the 72 Arm Neoverse-V2 cores paired with HBM3 memory. With this GPTshop.ai GH200 system actually being in workstation form, I also ran some additional benchmarks looking at the CPU capabilities of the GH200 compared to AMD Ryzen Threadripper 7000 series workstations.

https://www.phoronix.com/review/nvidia-gh200-amd-threadripper

[coder]

Thanks for more benchmarks, Michael & GPTshop.ai !

One question: do the compile benchmarks compile for the native platform, in each case? From a practical standpoint of how users are likely to do builds, it'd make sense if they did. However, from a purist benchmarking point of view, I'd suggest a common target (probably x86) be selected and cross-compilation be performed on non-x86 platforms.

Although it might sound like it should be slower, there's no intrinsic reason why cross-compilation should put non-native platforms at a disadvantage. Rather, it would make such benchmarks truly apples-to-apples.

Oh, and a second question: were any doubts of SSD thermal throttling adequately put to rest? That could also affect things like the compilation benchmarks.

[coder]

GPTshop.ai just asking out of curiosity: do you have any plans to offer a 2x Grace version, without Hopper?

[Nille]

How is the Power consumption over all tests?

[GPTshop.ai]

GPTshop.ai just asking out of curiosity: do you have any plans to offer a 2x Grace version, without Hopper?

I thought about it, but for now I focus on Grace Hopper. Maybe in the future.

[GPTshop.ai]

How is the Power consumption over all tests?

I assume so. Please ask Michael Larabel.

[Dawn]

The article implies in a bunch of places that Grace is using HBM3, but Nvidia's documentation would seem to suggest that the only memory available to the CPU side is the CPU's own memory (ie, LPDDR, not HBM - and not significantly more bandwidth than Epyc has.) Even if the CPU has access to GPU memory regions, which seems unclear (at least to me), it would be limited to only a relatively small portion of the HBM's bandwidth due to NVlink limitations.

Michael - have you run STREAM on the CPU? If it pulls <600GB/s, that's strongly suggestive that there is no use of the HBM no the CPU side.

[JEBjames]

Michael

typo on page 3

"performnace-per-dollar basis" should be "performance-per-dollar basis"

[ChirpinBird99]

Why is the 7980x at 7.79ghz?

[GPTshop.ai]

Oh, and a second question: were any doubts of SSD thermal throttling adequately put to rest? That could also affect things like the compilation benchmarks.

IHMO there should have been no SSD thermal throttling because there is a 200mm fan blowing fresh air directly onto it. But we will investigate this potential issue.