Nvidia Grace Hopper Superchip unveiled for AI-centric workloads
While Nvidia didn't have anything for gamers at its Computex 2023 keynote, its presentation was chock full of AI-centric products. Among them was the Grace Hopper Superchip, a combination of a Grace CPU and Hopper GPU. It will power Nvidia's DGX GH2000 supercomputer and a plethora of other services.
The Grace CPU packs 72 Arm Neoverse V2 CPU cores, up to 480 GB of LPDDR5X RAM (512 GB/s memory bandwidth) and a massive 117 MB of L3 cache. Its accompanying Hopper GPU can be configured with up to 96 GB of HBM3 memory (up to 4 TB/s memory bandwidth). Nvidia claims it offers up to 67 TFLOPS of FP32 performance. Similarly, its included Tensor cores can do 67 TFLOPS of FP64.
Combined, the Grace Hopper Superchip can guzzle as much as 1,000 Watts of power, depending on the configuration. The CPU and GPU are connected via an Nvlink module which operates at 900 GB/s, 5x faster than PCIe-Gen 5. Memory coherency allows both CPU and GPU threads to access each others' memory, allowing its new Transformer Engine trains AI models 30x faster than existing GPUs. Compared to x86-based solutions, the Grace Hopper Superchip is anywhere between 2.75x to 3.9x faster.
Devices powered by the Grace Hopper Superchip are already out and about in many regions including the United States and Europe. If recent rumours are accurate, we could even see a cut-down version of it alongside the next Nintendo Switch console.