Learn about cutting-edge optical interconnect technology for AI/ML clusters in this technical presentation from Broadcom and ByteDance experts. Explore how Co-Packaged Optics (CPO) technology addresses the demanding bandwidth, power, cost, and latency requirements of large-scale GPU clusters. Discover the manufacturing innovations and processes behind Broadcom's 51.2T CPO Switch, and understand how ByteDance implements this technology to meet AI training and inference demands. Examine performance statistics of the TH5-Bailly 51.2T CPO, including its ability to achieve 200G/lane speeds while reducing power consumption by over 40% and latency by over 70%. Gain insights into proposed standardization efforts for CPO operations monitoring aimed at strengthening the ecosystem for CPO-based networks.
Overview
Syllabus
AI/Ml Clusters Enabled By A 51.2 Tbps Cpo Switch
Taught by
Open Compute Project
Related Courses
-
Co-Packaged Optics (CPO) Integration and Enhancement with SONiC Switches
-
AI Infrastructure and Operations Fundamentals
5.0 -
Energy-Efficient High Bandwidth Optical Interconnect Solutions for Generative AI
-
Optical Interconnects for Large-Scale AI Clusters - A Meta Perspective
-
High-Performance Ethernet NIC for AI/ML - Thor 2 Architecture and Features
-
Insights from Production Scheduled Ethernet Fabric in Large AI Training Clusters
