The chip will be called Grace, after Grace Hopper, and is targeting a 10x speed advantage compared to x86-based systems.
Initial customers, says Nvidia CEO Jen-Hsun Huang, will be the Swiss National Supercomputing Centre (CSCS) and Los Alamos National Laboratory who will buy supercomputers based on Grace built by HPE’s Cray group for delivery in 2023.
Nvidia says that a Grace-based computer using Nvidia technology will reduce the amount of time it takes to train the natural language AI model GPT-3 on the 2.8 AI-exaflops Selene from fourteen days to two.
A number of chip companies including Qualcomm and AMD have tried, and failed, to get Arm-based CPUs into the server market.