Last month, Nvidia launched its latest 144-core Grace CPU Superchip at GTC, its first Arm-based CPU chip designed for data centers. Nvidia has not given the performance standard of this CPU before, but it has mentioned that EPYC is 1.5 times stronger than AMD, but it said nothing.
Join tip3x on Telegram
Moreover, the PPT of Ian Buck’s GTC speech, vice president of Nvidia’s accelerated computing business unit, has now been revealed, and sources found a benchmark comparison between Grace and Intel Ice Lake.
Nvidia claims that Grace is 2x faster and 2.3x more energy-efficient than Intel’s current Ice Lake in weather research and forecasting (WRF) models commonly used in HPC. It’s 1.5x faster than AMD’s previous-generation 64-core EPYC Rome 7742 processor in the SPECrate_2017 benchmark.
Furthermore, it will deliver twice the power efficiency of today’s AMD server chips when it ships in early 2023 when AMD already has Faster EPYC Milan, and probably no match for AMD’s EPYC Genoa released next year. In addition, Intel will also launch its Sapphire Rapids processors by 2023, and is not expected to be too far behind this Grace.
Besides, Nvidia Grace CPU Superchip is a processor based on the Arm v9 architecture, with a total of 144 cores in the two chips, which supports Nvidia’s NVLink-C2C interconnect technology, and can provide a throughput of 900 GB/s.
In addition, the chip uses 1TB of LPDDR5x ECC memory, which can provide up to 1TB/s of memory bandwidth, double that of other data center processors that support DDR5 memory. Therefore, this processor benefits from this powerful memory throughput and has considerable advantages in the above-mentioned WRF weather forecast model.
Nvidia says its simulations of the 144-core Grace chip show it will be 2x faster and 2.3x more power-efficient than two 36-core 72-thread Intel Ice Lake Xeon Platinum 8360Y processors in weather research and forecasting.
