As many of our readers will know NVIDIA Grace is a high performance CPU for servers and data centers based on ARM, and which can be purchased in CPU and CPU + GPU configurations. The latter gives shape to the GH200 superchip that we recently presented to you.
NVIDIA Grace uses the architecture Neoverse N2, and can use LPDDR5x memory with support for ECC technology, short for error correction. Its most important features also include its configuration with up to 144 cores, its full compatibility with the NVIDIA software ecosystem, and its excellent performance per watt consumed.
Just like that statement is generic, so NVIDIA wanted to show how far the performance per watt ratio can go in the Grace CPU. In the attached graph we can see a performance comparison that the green giant has shared taking advantage of the Hot Chips 2023 scenario, where it has faced said CPU with Intel Sapphire Rapids and AMD EPYC Genoa processors.
It is important that you keep in mind that this performance comparison has been carried out performance tests adjusting all processors to the same level of consumption, which allows us to see the performance value per watt consumed that each of them offers. The results are in the graph on the right, which is labeled “5MW Data Center Throughput.”
In this scenario, which represents a consumption of 5 megawattsNVIDIA Grace comes to offer up to 250% more performance and achieves a minimal improvement, in the worst case, of 70%. The results obtained by this NVIDIA super CPU are very positive, there is no doubt about that, and they demonstrate the potential of ARM architecture in data centers.
In the tests, dual-socket configurations of AMD EPYC Genoa and Intel Sapphire Rapids were used, which means that in the first case we have two AMD EPYC 9654 processors with 96 cores and 192 threads each, and in the second we have two Xeon Platinum 8480 with 56 cores and 112 threads each. Doing numbers we get a total of 192 cores and 384 threads in the first case and 112 cores and 224 threads in the second.
