Inside ThetaGPU and Perlmutter's NVidia Ampere A100 GPU

Introduction

This webinar presented by Max Katz from Nvidia is part of the ALCF Developer Sessions, also open to NERSC users. 

Date and Time: 9 am - 10 am (Pacific time), July 28 (Wednesday), 2021

Abstract

The NVIDIA Ampere A100 GPU, which powers the ALCF's ThetaGPU system and NERSC's Perlmutter system, builds on the excellent performance of the NVIDIA V100 GPU and includes several new features for both HPC and deep learning. Particularly, tensor cores have been improved (including FP64 compute and efficient evaluation of sparse computation); the L2 cache is significantly expanded and includes new residency controls; the SMs have enhanced support for asynchronous operations involving shared memory; and the new Multi-Instance GPU feature allows true isolation of partitions of the GPU. NVIDIA has also invested significant effort in ensuring that both HPC applications and DL frameworks achieve excellent performance on A100. This talk will describe these (and other) aspects of A100 so that computational scientists can get a better idea of what is possible on this architecture. 

Please find more information about the speaker and register at the event page of the ALCF Developer Sessions. 

Presentation Materials

• Slides
• Video Recording