So far, Nvidia supports only x86 and IBM’s Power processors, however, things could change from the end of 2019 to run its own software such as libraries and frameworks on ARM CPUs. First supercomputers are in the works.
Nvidia announced at the ISC (International Supercomputing Conference) in Frankfurt, to open its own software stack for ARM supercomputers. This includes Cuda X, HPC libraries and GPU-accelerated artificial intelligence frameworks, as well as compilers and profilers to take full advantage of the manufacturer’s graphics chips.
So far, the High-Performance Computing (HPC) segment is dominated primarily by Intel x86-based systems, followed by supercomputers with IBM’s Power models; AMD has been barely represented for years. A small percentage uses in-house developments and ARM computers are almost irrelevant so far.
NVIDIA has produced 22 of the 25 most efficient supercomputers in the world, according to the updated Green500 list today, thanks to CUDA GPUs that can take care of many intensive jobs freeing up the CPUs: GPUs are more specialized and have higher efficiency.
If it goes by the quotations of Nvidia’s partners, but this could change in the coming months: the CPU developers Ampere and Marvell, the system manufacturer Atos, Cray and HPE and supercomputer centers from Germany, Finland and Japan have pledged their support. Among others, the Rikagaku Kenkyujo (Riken) research institute is working on Fugaku, a Japanese supercomputer that uses ARM-based A64FX processors developed by Fujitsu.
Nvidia has a lot of experience with ARM designs for many years, as its own Tegra series relies on this RISC architecture: The Tegra APX 2500 was released in 2008 since the Tegra K1 of 2014 Nvidia also uses its own ARM cores like Denver and Carmel was developed for the Tegra Xavier. In addition to ARM cores stuck in the Tegras and Geforce GPUs, based on architectures such as Kepler, Maxwell, Pascal, and Volta. The Tegra X1 is used by Nintendo for the handheld console switch.