NVIDIA Unveils Grace CPU Superchip and Blazing Fast H100 Datacenter GPU
NVIDIA reveals its powerful new Grace CPU Superchip featuring 144 ARM cores and introduces the Hopper-based H100 datacenter GPU that delivers up to six times the performance of its predecessor.
NVIDIA officially unveils its Grace CPU Superchip, a powerful ARM-based server processor that features two interconnected Grace CPUs. The chip integrates a staggering 144 ARM cores into a single socket and utilizes built-in LPDDR5x memory to deliver bandwidth speeds of up to 1 terabyte per second. This design functions similarly to Apple's M1 Ultra, connecting two processors seamlessly to boost overall performance for demanding datacenter workloads.
Alongside the new processor, NVIDIA introduces its highly anticipated Hopper architecture and the H100 datacenter GPU. Built using an 80 billion transistor design, the H100 delivers up to six times the performance of the company's previous Ampere-based generation. Clients have the option to configure the Grace CPU Superchip as a standalone system or pair it with up to eight of these new Hopper-based GPUs to maximize computing power.
NVIDIA claims the Grace CPU Superchip operates twice as fast as traditional servers and expects it to achieve approximately 740 points in SPECrate®2017_int_base benchmarks. The company plans to begin shipping the new CPU to clients, including the Department of Energy, in the first half of 2023. While mainstream consumer GPUs are not part of this announcement, the new Hopper architecture clearly signals a massive leap forward for enterprise and artificial intelligence computing.