We propose an HPC network architecture with co-packaged optics enabling 128-port 51.2-Tb/s switches. Simulations for a >34,000-GPU system show up to 11.2x throughput improvement over a Summit-like supercomputer, opening the way to direct-network-attached GPUs.
© 2020 The Author(s)
You do not have subscription access to this journal. Citation lists with outbound citation links are available to subscribers only. You may subscribe either as an OSA member, or as an authorized user of your institution.
Contact your librarian or system administrator
Login to access OSA Member Subscription