NVIDIA’s Tesla P100 GPUs will power Japan’s fastest AI supercomputer

The Tokyo Institute of Technology has announced plans for a new AI supercomputer called Tsubame3.0. The new system will use NVIDIA’s Tesla P100 GPUs to double performance over its predecessor, Tsubame2.5.

The new NVIDIA Pascal-based GPUs are three times as efficient as their predecessors and can deliver up to 12.2 petaflops of double precision performance. This would place the Tsubame3.0 in the top 10 of the world’s fastest supercomputers, with up to 47 petaflops of what NVIDIA calls AI computation.

This really refers to half precision computation, or a 16-bit floating point format, which is one of the precision sweet spots for AI computation.

Tsubame3.0 could also work in conjunction with Tsubame2.5, where the two supercomputers would deliver up to 64.3 petaflops and become Japan’s top performing AI supercomputer.

The supercomputer will be used primarily for education and technology research at the Tokyo Institute of Technology, but it will also be open to researchers from other universities or the private sector.

The high-performance computing (HPC) segment has traditionally dabbled in things like climate modeling and scientific simulations, but the growing prevalence of AI-optimized chips has blurred the line between the HPC and machine learning markets.

The growing efficiency and performance of the latter chips has made it more attractive for supercomputers to adopt them, and we may see wider uptake of machine learning chips in next-generation supercomputers, especially as many simulations involve things like image recognition that could benefit from machine learning techniques.

Source: NVIDIA

Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.