The latest benchmarks of NVIDIA's Grace CPU reveal twice the performance compared to the fastest x86 CPUs from AMD and Intel at the same power consumption level.

Recently, NVIDIA also announced the most powerful GPU for AI and computing, known as GH200, accompanied by the world's fastest HBM3e memory. It will be paired with NVIDIA's latest CPU to form a complete system, completely eliminating the need for Intel or AMD CPUs.
Some key highlights of Grace include:
- Unleash the power of high-performance computing and cloud computing with NVIDIA's revolutionary Grace CPU. Featuring a super chip with up to 144 Arm v9 CPU cores, the world's first LPDDR5x memory with ECC, delivering an astonishing total bandwidth of 1TB/s.
At Hot Chips 2023, NVIDIA's Head of Science, Bill Dally, presented performance comparisons between the NVIDIA Grace Superchip and competing dual-socket x86 solutions. These include AMD's EPYC 9654 with 96 cores & 192 threads and Intel Xeon Platinum 8480+ with 56 cores and 112 threads. Each motherboard can accommodate two of the mentioned CPUs, resulting in a total of 192 cores for AMD and 112 cores for Intel platforms.

Meanwhile, the NVIDIA Grace CPU offers a total of 144 cores, supports LPDDR5X memory up to 960GB with a bandwidth of up to 1TB/s, and consumes a total power of 500W. Additional technical specifications include 117MB of L3 cache and 58 Gen5 lanes, all utilizing the TSMC 4N processing node.
NVIDIA has selected various server applications as benchmarks, including Weather WRF, MD CP2K, Climate NEMO, CFD OpenFOAM, and Graph Analytics GapBS BFS. Across all benchmarks, NVIDIA's Grace Superchip provides up to 40% better performance than AMD's Genoa CPU while significantly outperforming Intel's Sapphire Rapids CPU. Most benchmarks are on par with AMD Genoa, whereas the two AMD CPUs consume up to 640W, while the NVIDIA Grace Superchip runs at 500W.
The performance comparison becomes even more intriguing when comparing against a real-world large-scale data center application. Benchmarking a 5 MW data center throughput reveals that NVIDIA's Grace superchip can deliver performance gains of up to 2.5 times. Therefore, the Grace CPU could be a game-changer akin to NVIDIA's Tensor Core GPUs, which have dominated the HPC and AI space.
Read more: NVIDIA unveils the 'ultimate' GPU superchip, priced at nearly a billion dong, not accessible to everyone.