Nvidia’s Grace CPU doubles its AMD and Intel rivals

Nvidia has shared new benchmarks for its Arm-based Grace GPU, which will power next-generation data centers and servers. The company’s Grace CPU solution is shown to offer twice the performance at the same power consumption when compared to AMD Genoa and Intel Sapphire Rapids x86 chips.

Nvidia’s Grace CPU knows no rivals

Powered by Arm Neoverse N2 cores, the Grace CPU will be used in Nvidia’s Superchips with both CPU+CPU and CPU+GPU versions. The company had showcased the GH200 GPU, which has the world’s fastest HBM3e memory, in the past months and positioned it in the Grace Hopper Superchip solution. Another pillar of this platform will be the Grace CPU, as the name suggests.

Some of Grace’s highlights are as follows:

• High performance CPU for HPC and cloud computing
• Superchip design with up to 144 Arm v9 CPU cores
• World’s first LPDDR5x with ECC memory and 1TB/s total bandwidth
• 900GB/s compatible interface, 7x faster than PCIe Gen 5
• Twice the packing density of DIMM-based solutions
• 2x more performance per watt than today’s leading CPU
• Support for all Nvidia software stacks and platforms, including RTX, HPC, AI and Omniverse

Sharing its details at Hot Chips 2023, Nvidia presented performance comparisons between Grace Superchip and its competitors’ dual-socket x86 solutions. These include the EPYC 9654, AMD’s fastest 96-core and 192-thread solution, and Intel’s flagship Xeon Platinum 8480+ with 56 cores and 112 threads. Since the solutions run on a dual-socket configuration, the total number of cores increases to 192 for AMD and 112 for Intel.

We know the features of the Grace CPU from the Grace Superchip configuration. Accordingly, Nvidia’s solution offers 144 (72 Arm Neoverse V2 per chip) cores, up to 1TB/s raw bandwidth, up to 960GB of LPDDR5X memory and 500W power consumption. Additional features include 117MB of L3 cache, 58 PCIe Gen5 lanes, and the TSMC 4N manufacturing process.

Intel is far behind

The benchmarks selected by Nvidia cover a wide range of server applications. According to benchmarks, Grace Superchip CPUs offer up to 40% better performance than AMD’s Genoa CPUs, while well ahead of Intel’s Sapphire Rapids CPUs. Most of the benchmarks were on par with the Genoa and even that is great for Grace because both of these chips have a total TDP of 640W (EPYC 9654 – 320 Watts) while the Grace Superchip is powered by 500W.

However, the solutions get even more interesting when compared to a true large-scale data center application. A 5MW Data Center throughput benchmark shown shows that Nvidia’s Grace Superchips are capable of delivering up to 2.5x the performance while being hugely efficient in the same benchmarks. For data center and server customers investing in these workloads, Grace CPUs could be a big game changer, just as Nvidia’s Tensor Core GPUs dominate the HPC and AI space. And finally, let’s mention that these comparisons are made by Nvidia.