twice the performance of AMD and Intel

twice the performance of amd and intel

Tom Henry

twice the performance of AMD and Intel

NVIDIA has shown new benchmarks of its “Super APU” NVIDIA Grace. We must remember that we are dealing with a superchip that makes use of an Arm processorbased on the Neoverse N2 cores, along with an NVIDIA GH200 GPUthe world’s most powerful graphics chip for AI and high computing workloads.

For you to have a context, the NVIDIA GH200 is accompanied by a CPU 144 cores Arm Neoverse N2 (Arv v9). This processor supports up to 960 GB of LPDDR5X memory with a gross bandwidth of up to 1TB/s. The CPU is accompanied by a graphics chip based on the Hopper architecture (4nm from TSMC). It boasts of having nothing less than 282 GB of HBM3e memory.

For context, this next-generation memory offers 3.5 times more memory capacity, and multiply the bandwidth by 3 the most advanced product currently available on the market. All results in AI performance of 8 petaflops.

NVIDIA Grace CPU boasts great potential


Now we come to the interesting part, and that is that this processor showed its worth compared to the most advanced x86 solutions on the market. On the one hand it was compared with a processor AMD EPYC Genoaand with a Intel Sapphire Rapids. Specifically, with a AMD EPYC 9654 with 96 cores and 192 threads, and a Intel Xeon Platinum 8480+ with 56 cores and 112 threads.

As if that were not enough, the benchmarks are linked to a Dual Socket configurationso we really talk about systems with two AMD CPUs and two Intel CPUs. Both processors are the best product of each company.

NVIDIA Grace CPU Superchi vs AMD EPYC vs Intel Xeon

When it comes to testing, there are many different tests to cover a wide spectrum of how these CPUs are used in servers. These are Weather WRF, MD CP2K, Climate NEMO, CFD OpenFOAM, & Graph Analytics GapBS BFS. In all cases, the NVIDIA Grace superchip offered up to 40% more performance compared to AMD EPYC processors, which are already much further ahead in performance than the Intel option. Now, while the configuration of the two AMD EPYC 9654 generated a consumption of 640Wthe NVIDIA solution stood out with 500W.

In a real world environment, the NVIDIA Superchip is up to 2.5 times faster


Outside of benchmarks, in a real application, inside a large-scale data center. Taking as a reference a data center of 5 megawatts of power. that was where NVIDIA Grace can offer up to twice the performance and, at the same time, be much more efficient in the same tests. For server and data center customers investing in these workloads, Grace CPUs can be a game changer. Just like NVIDIA Tensor Core GPUs have dominated the high performance computing space and artificial intelligence.

Tests show that, at the same power, Grace systems offer up to twice the performance of today’s x86 servers in various CPU workloads. In addition, Arm’s SystemReady program certifies that Grace systems will run existing Arm operating systems, containers, and applications without modification.

This should not surprise us, since at the beginning of the year, NVIDIA already announced that they had designed Grace to be 2x more efficient than the best CPUs from AMD and Intel.

Leave a Comment