Hemal Shah, Broadcom, and James Wynia, Dell Technologies | SC23
summary
Broadcom shows production silicon for large AI fabrics: Jericho3-AI (a 100 billion+ transistor supercomputer switch chip) alongside Tomahawk 5, a top-of-line ASIC offering 64×800G ports for multi-stage switch fabrics. Ultra Ethernet adds multipathing, adaptive routing, congestion control and selective RDMA retransmit to scale across clusters (vendors mention half‑million‑server scales). Dell supplies integrated reference architectures, cabling, optics and cluster management so deployed systems hit line‑rate GPU feeding (400G today, moving to 800G) and are field‑proven.
AI-written summary · about
chapters (11)
- 0:00Introduction to HPC
- 1:02Broadcom and Dell partnership
- 2:23Ultra ethernet technology
- 3:39Hardware showcase
- 5:29Differentiation and scalability
- 7:07Open standards in AI
- 9:14Overcoming network bottlenecks
- 11:11Business impact of AI
- 12:58Integrated solutions
- 16:06Ecosystem interoperability
- 17:13Future outlook and wrap-up
