AMD claims Nvidia's Grace CPU Superchip, Arm are no match for its Epyc Zen 4 cores

AMD has claimed its current datacenter silicon is already more than twice as fast, and up to 2.75 times more efficient, than Nvidia's Grace CPU Superchips. The chip design firm's assertions came after its own testing, published last week, in which it considered Nvidia's 2022 Grace CPU Superchip. That product combines a pair …

[HuBo]

Gong!

Cool benchmarks! It is notable that AMD provides a lot more details about where their numbers come from than can be found in the Nvidia datasheet and blogs. They note for example (among others) that, for Quantum ESPRESSO, on ARM, "scalapack is not available in this library" (i.e. challenges remain in the benchmarking software set on ARM).

The perf./Watt results are possibly the most interesting. In my mind, they confirm Green500 where, at 52.9 GF/W, 2022's AMD EPYC+MI250x 1.2 EF/s Frontier still beats 2024's brand new Nvidia GH200 Alps' 52.0 GF/W (for 0.27 EF/s). As Tobias wrote, it'll be fascinating to see the "head-to-head between the GH200 and AMD's MI300A", whenever these pugilists are ready!

[talk_is_cheap]

They are both fast, now what problem are you trying to solve.

These system designs have to be mapped to a business problem before any choice can be made, the benchmarks just tell a designer they are nice and fast. If the business problem needs all cores to work on the same memory-based data set you are going to start looking at the Nvidia solution first as the memory I/O speed and processor interconnect are designed for just that with performance numbers beyond anything AMD or Intel can offer. On the other hand, if the problem does not need a unified memory model or needs more than 960GB of RAM the Nvidia solution may not fit so well.

Note - AMD gave Nivida an advantage in the benchmarks as the 480GB solution is reported to have a memory bandwidth of 1024GB/s, while the larger 960GB solution is listed at just 768GB/s.