Hasan Siraj, Broadcom | SC23
summary
Broadcom argues that modern AI training makes the network the “computer,” backing that with products and numbers: Jericho3 and Tomahawk 5 target congestion and scaling, Jericho shows 12% better performance than InfiniBand, supports a 32,000‑node domain and claims sub‑10 nanosecond failover, and 51.2 TB switches appear from multiple vendors. Ethernet is framed as cost‑effective and 50 years young, a 10% network speedup pays for itself, and a 10‑member Ultra Ethernet Consortium aims to standardize RDMA, multipathing and selective retransmit.
AI-written summary · about
chapters (11)
- 0:00Welcome to SC23
- 1:03Networking's critical role in AI
- 3:13AI workload challenges
- 5:19Hyperscalers and AI deployments
- 7:27Optimizing GPUs vs networking
- 9:12Ecosystem and best practices
- 10:59Jericho fabric performance
- 12:47Architectural flexibility
- 13:32VMware and networking revival
- 14:59Ultra Ethernet Consortium
- 17:33Open standards for AI
