Live Panel: Scale-Out Networking for AI Supercomputers
summary
Panel describes how real supercomputers tie tens of thousands of accelerators into one fabric and reports implementing MRC across vendors to get synchronized training working, uses RDMA and Rocky to push utilization toward near 100%, and shows concrete problems: observability gaps and gray failures that force homegrown telemetry and agentic chatops for maintenance. It warns scale will exceed single data centers, raises KV cache network load and long‑distance links as the next hard limits.
AI-written summary · about
chapters (10)
- 0:00Introduction and panel scope
- 1:20Panelist introductions
- 5:56Defining networking challenges
- 8:30Operations and observability
- 12:52Integrating AI into operations
- 15:23Solution development strategies
- 19:15Future architecture trends
- 21:31Converging networking domains
- 23:51Managing diverse environments
- 25:50Final advice and conclusion
related talks
21mDesigning Scalable Networks for Large AI Clusters: Challenges and Key Insights
06mOpening Remarks
17mMeta’s Network Journey to Enable AI
19mAI/ML Networking Challenges: The Fast and the Finicky!
21mFuture Challenges for HPC Networking @Scale
35mPanel: Kubernetes Networking: Present and Future
07mWelcome Remarks
31mLive Q+A 1