NVIDIA expands supercomputing infrastructure to support Arm-based processor architecture.

Nvidia announced that its CUDA-X AI and HPC libraries, graphics processing unit-accelerated AI frameworks, and software development tools will support Arm-based machines by the end of the year. Arm-based supercomputers should enable far greater scale thanks to their greater power-efficiency, according to the company. Nvidia Chief Executive Jensen Huang stated, "As traditional compute scaling ends, power will limit all supercomputers. The combination of Nvidia's CUDA-accelerated computing and Arm's energy-efficient CPU architecture will give the HPC community a boost to exascale," a term referring to computers capable of performing a quintillion operations per second.

The decision to support Arm CPUs came about because of broad and growing industry interest, according to Ian Buck, Nvidia's general manager and vice president of accelerated computing. "What makes Arm interesting is it's very open," Buck said. "[It provides] flexibility for new ways to connect CPUs and GPUs and do more energy-efficient computing." Nvidia's infrastructure is already used in 22 of the world's top 25 most energy-efficient supercomputers, supporting x86 and POWER-based chips. The addition of Arm support aims to expand Nvidia's presence in the HPC space to support even more advanced AI workloads.

Nvidia unveiled the DGX SuperPOD, described as the world's 22nd fastest supercomputer, along with a reference architecture for companies seeking to deploy it in their own or outside data centers. Designed to provide AI training infrastructure for deploying massive fleets of autonomous vehicles, the system was built in just three weeks and comprises 96 of Nvidia's DGX-2H supercomputers integrated using new data center interconnect technology acquired through the company's purchase of Mellanox earlier this year.

The DGX SuperPOD delivers 9.4 petaflops, or quadrillion floating-point operations per second, and is designed to train neural networks for self-driving cars so they understand the "rules of the road." This performance level can reduce training time on the popular image classification ResNet-50 AI algorithm from 25 days to less than two minutes. Clement Farabet, vice president of AI infrastructure at Nvidia, stated, "AI leadership demands leadership in compute infrastructure. Few AI challenges are as demanding as training autonomous vehicles, which requires retraining neural networks tens of thousands of times to meet extreme accuracy needs."