Abstract
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)
PDF ArticleMore Like This
D. Kuchta, M. Meghelli, P. Pepeljugoski, L. Schares, M. Schultz, P. Maniotis, P. Fortier, E. Tucotte, C. Bureau, M. Pion, Y. Cossette, G. Jutras, B. Sow, B. Parikh, S. Ostrander, S. Li, D. Becker, F. Gholami, H. Bagheri, F. Flens, G. Light, and B. Wang
NeTu1D.2 Photonic Networks and Devices (Networks) 2022
P. Maniotis, L. Schares, and D. M. Kuchta
Th1D.3 Optical Fiber Communication Conference (OFC) 2023
Shang Yu, Bingli Guo, Wenzhe Li, Yu Zhou, Xin Li, and Shanguo Huang
W4I.5 Optical Fiber Communication Conference (OFC) 2018