NVIDIA Grace loses to Intel Sapphire Rapids in the first HPC benchmarks
NVIDIA is better known as a GPU company, but it also hopes to become a CPU company.
The company has an Arm-based server processor known as Grace, which it sells in configurations with either two Grace processors or with a Hopper GPU, both of which are intended for server tasks. However, NVIDIA faces an uphill battle in this category as it competes with Intel, AMD and other Arm-based server chips. This is reported by Extreme Tech.
The first published High Performance Computing (HPC) benchmarks for Grace come from several supercomputing labs that measure CPU performance.
The Barcelona Supercomputing Center and the State University of New York tested Grace in both dual CPU and CPU + GPU combinations, known as Grace-Grace and Grace-Hopper. NVIDIA hardware has been tested on x86 server processors and the results are impressive.
For reference. The NVIDIA Grace processor has 72 Arm cores with 480 GB of LPDDR5X memory. This makes it a unique processor as it offers a large core count with two Grace chips (by default) offering 144 cores and also uses mobile memory (with ECC) instead of HBM like most server CPUs.
Benchmark tests include a wide range of tests. TL;DR NVIDIA Grace products are competitive against older architectures like Intel Skylake and AMD Milan. At the same time, they had a significant cost and thermal advantage thanks to the Arm architecture.
However, the NVIDIA Grace CPUs were outperformed by the latest Intel Sapphire Rapids CPUs in several HPC benchmarks. For example, as Tom’s Hardware points out, compared to a pair of 48-core processors, NVIDIA’s Intel Grace Xeon Max 9468 was faster in only three of eight tests.
Overall, the results for Grace are relatively impressive compared to existing platforms, but power consumption data was not included in the tests.
NVIDIA’s dual-socket Grace processor has a TDP of 500W, which compares to 700W for the dual-socket Sapphire Rapids chip. Therefore, NVIDIA has some leadership in this segment, which is a critical factor in the deployment of data centers.
Moreover, none of these benchmarks include the latest AMD Zen 4 server chips, as well as Epyc Genoa processors.
With AMD starting to take server market share away from NVIDIA, many data center operators will be looking closely at both AMD and NVIDIA for the foreseeable future.
Read also on ProIT: NVIDIA is preparing the Chinese AI chip H20 to compete with Huawei.
Subscribe to ProIT on Telegram so you don’t miss a post!