Jas Tremblay, Broadcom, and Ihab Tarazi, Dell Infrastructure Group | SC23
summary
Dell and Broadcom describe concrete outcomes: Dell’s XE9680 multi‑GPU server (eight GPUs + memory fabric) delivers about 10× single‑GPU performance and ships today, while AI servers run roughly $200–$250k versus $5–$10k for typical servers. Networking not GPUs becomes the cluster bottleneck; work focuses on matching link speeds to compute, optimizing NIC count, automation, and driving adoption of PCIe with CXL inside servers, plus an open non‑NDA fabric roadmap to speed standards and interoperability.
AI-written summary · about
chapters (12)
- 0:00Welcome and introduction
- 0:42AI server history and evolution
- 2:28AI ecosystem challenges and opportunities
- 5:40Compute vs AI server standardization
- 8:16Importance of standards in AI servers
- 10:24Cluster setup complexity and bottlenecks
- 11:51Standardization goals and timeline
- 13:34Co-opetition and industry collaboration
- 15:41Cost and scalability of AI servers
- 17:53Future of AI server innovation
- 20:02Next year's Supercomputing predictions
- 20:46Closing remarks and thanks
