Abstract
We investigate the viability of optically switched network for ML accelerator clusters and compare it to a leaf-spine network with 256/1024 GPUs. Results show almost ideal throughput, sub-µs latency and zero packet-loss for ¡0.6 traffic-load.
© 2023 The Author(s)
PDF Article | Presentation VideoMore Like This
Huaxi Gu, Xiaoshan Yu, Yunfeng Lu, Hong Zou, and Shuo Li
Th3D.1 Optical Fiber Communication Conference (OFC) 2023
Xuwei Xue, Bitao Pan, Sai Chen, Kristif Prifti, Xiaotao Guo, Fulong Yan, Rafael Kraemer, Shaojuan Zhang, Chongjin Xie, and Nicola Calabretta
Tu5J.3 Optical Fiber Communication Conference (OFC) 2021
Wang Miao, Fulong Yan, Oded Raz, and Nicola Calabretta
W1J.3 Optical Fiber Communication Conference (OFC) 2016