Watch a technical presentation from ByteDance and Broadcom architects exploring how Scheduled Ethernet Fabric technology optimizes large-scale AI training clusters. Discover the architecture behind connecting tens of thousands of GPUs efficiently, with detailed insights into achieving extensive GPU scale-out, managing diverse parallel workloads through multi-tenancy, and implementing resilient networking against failures. Learn from ByteDance's real-world benchmarking results and deployment experiences with this fabric technology, while gaining perspective on the importance of open ecosystems for continued innovation in AI infrastructure. Understand key requirements for high-performance network fabrics that maximize computational power across massive GPU clusters handling various AI workloads.
Overview
Syllabus
Scheduled Ethernet Fabric for Large scale AI training cluster
Taught by
Open Compute Project
Related Courses
-
Insights from Production Scheduled Ethernet Fabric in Large AI Training Clusters
-
Next Generation Fabric for AI/ML Use Cases - Disaggregated Scheduled Fabric Using SONiC
-
Congestion Management in Ethernet-Based Networks for AI Cluster Fabric
-
High-Performance Ethernet NIC for AI/ML - Thor 2 Architecture and Features
-
Orchestration Needs for AI Clusters at Scale - Lessons Learned from Two Leading Providers
-
Exploring the Wilderness - Optimizing Ethernet Fabrics for AI Workloads
