NVIDIA GH200 CPU Performance Benchmarks Against EPYC Zen 4 & Xeon Emerald Rapids

[phoronix]

Phoronix: NVIDIA GH200 CPU Performance Benchmarks Against EPYC Zen 4 & Intel Xeon Emerald Rapids

Kicking off our NVIDIA GH200 Grace Hopper benchmarking at Phoronix is an initial look at the 72-core Grace CPU performance with 96GB of HBM3 memory. Here are some initial benchmarks of the Grace CPU performance while the Hopper GPU benchmarks will be coming in a follow-up article.

https://www.phoronix.com/review/nvidia-gh200-gptshop-benchmark

[spiral_23]

and what said linus?

[sophisticles]

The 41 grand is for the base model and it goes up from there.

They offer some beautiful glass cases for these systems, almost like they read my posts to System76 telling them to do the same thing.

They make some amazing performance claims:

Its performance in every regard is almost unreal (up to 284 times faster than x86).

Compared to 8x Nvidia H100, GH200 costs 5x less, consumes 10x less energy and is not far off in terms of performance.

The H100 costs 30 grand each, so 8 of them would cost 240 grand, divided by 5 equals 48 grand, so that claim is accurate.

Assuming the rest of their claims are accurate:

One H100 consumes 700 Watts of power, eight would consume 5600 Watts of power, meaning that this system consumes roughly 560 Watts of power.

This seams like very low power consumption.

I can't wait for NVIDIA's desktop ARM chips in 2025.

[varikonniemi]

Interesting it can both lead and trail the benchmarks. This would lead me to think it is extremely unbalanced and tries to roll on some magic tricks they decided to employ to show certain benchmarks can be dominated.

[Michael]

Interesting it can both lead and trail the benchmarks. This would lead me to think it is extremely unbalanced and tries to roll on some magic tricks they decided to employ to show certain benchmarks can be dominated.

Not necessarily tricks but how well a given software package is optimized for AArch64. As well as if its memory bandwidth intensive stands to benefit the most from HBM, etc.

[varikonniemi]

Not necessarily tricks but how well a given software package is optimized for AArch64. As well as if its memory bandwidth intensive stands to benefit the most from HBM, etc.

Have you ever seen a new product win some tests, and trail others in some tests? This seems to me as evidence they targeted their silicon to a very narrow and modern use-case, which led to very bad performance in some other tests that did not fit this optimizations.

[Michael]

Have you ever seen a new product win some tests, and trail others in some tests? This seems to me as evidence they targeted their silicon to a very narrow and modern use-case, which led to very bad performance in some other tests that did not fit this optimizations.

Xeon Max?

[Brane215]

How much do these thingies cost ?
I0ve found somewhere price estimate $25k-$40k for H100, which is I assume previous-gen.

THis seems outrageous for H200. Am I missing something ?
Anyone with a link with a pricetag ?

[Michael]

How much do these thingies cost ?
I0ve found somewhere price estimate $25k-$40k for H100, which is I assume previous-gen.

THis seems outrageous for H200. Am I missing something ?
Anyone with a link with a pricetag ?

Just the system prices, as noted in the article.

[stormcrow]

Xeon Max?

You're both correct. These systems are narrowly tailored for specific tasks, but it's not unheard of for supposedly general purpose systems to perform really poorly with some loads while being well suited for others. It's mostly a matter of what they're architected and tuned to do.

There's No Such Thing as a General Purpose Processor

Point is, regardless of the processor, there are going to be somethings it's good at. Other tasks it tanks on because there's a near infinite number of tasks a processor can be set to work on, but only a small subset will it be really good at. Some good examples are NUMA based domain computers (some Xeon systems and others) where the processors can only quickly access the physical RAM it's physically adjacent to but hits a massive performance wall when it needs to go outside its assigned NUMA domain. Sun's Niagra SPARC architecture which was designed for highly concurrent loads like transactional databases but was awful at single threaded computation, and GPUs which are excellent at massively parallelized vector math and adjacent math algorithms but not much else.