Nvidia Builds In-House GPU Supercomputer

Matthew Broersma,
Image credit: Nvidia

Selene supercomputer uses new A100 GPUs and Mellanox InfiniBand interconnects to deliver 1 exaflop, putting it in top 10 performance rankings worldwide

Nvidia said manufacturers are announcing more than 50 servers using the company’s new A100 graphics processing units (GPU) in areas including AI, data science and scientific computing.

The company said it has even built an in-house supercomputer called Selene based on A100 GPUs.

Nvidia announced the A100 in May using the SXM form factor, which uses Nvidia’s own NVLink interconnect.

At the ISC High Performance 2020 online high-performance computing event, Nvidia announced a PCIe version of the GPU that’s compatible with standard rack-mounted servers.

Image credit: Nvidia
Nvidia’s Mellanox InfiniBand interconnect technology. Image credit: Nvidia

GPU supercomputing

Based on the new Ampere architecture, the chip features a significant performance jump over the previous Volta architecture, and can be partitioned into seven separate accelerators for specialised applications.

The performance improvement means that GPU-based servers taking up only four racks and costing $3 million (£2.4m) can deliver the same processing power as servers based on conventional CPUs that take up 22 racks and cost $20m, Nvidia said.

The manufacturers delivering the new A100-based servers include Asus, Cisco, Dell, Fujitsu, Hewlett Packard Enterprise, Lenovo and Supermicro.

Thirty of the systems are due this summer, with more than 20 more expected by the end of 2020, Nvidia said.

“The sheer breadth of Nvidia A100 servers coming from our partners ensures that customers can choose the very best options to accelerate their data centers for high utilisation and low total cost of ownership,” said Ian Buck, Nvidia’s vice president of accelerated computing.

The company said it combined its GPUs with InfiniBand technology from Mellanox, which it acquired for $7bn last year, to create a 1-exaflop supercomputer called Selene that it said ranks in the top 10 of supercomputing systems worldwide.

Image credit: Nvidia
Nvidia’s PCIe A100 GPU. Image credit: Nvidia

Weather forecasting

The system took less than a month to design, Nvidia said.

ISC said that Nvidia’s GPUs were already used in 125 of the world’s top 500 supercomputers as of last year, with the figure rising to more than 300 for those including either Nvidia GPUs or Mellanox interconnects.

GPUs have become popular in high-performance computing as they help accelerate specialised workloads such as artificial intelligence, data anlalytics and big data visualisation.

A number of weather-forecast supercomputers around the world use Nvidia-based supercomputers, the company said, including the Spanish Meteorological Agency, the China Meteorological Administration, the Finnish Meteorological Institute, NASA and the Royal Netherlands Meteorological Institute.

It said the Beijing Metorological Service is using 200 Gigabit HDR InfiniBand interconnects for a new supercomputing platform it plans to deploy for the 2022 Winter Olympics in Beijing.