Microsoft HotChips presentation reveals bandwidth wall in AI cluster architecture
bookwormengr · x · 2026-08-31
At HotChips, a Microsoft presentation highlighted a "bandwidth wall" in AI cluster architecture: bandwidth between local accelerators is 6-7 400G (non-switched), while between different compute trays/racks it drops to 2 400G (switched). Although Microsoft uses a unified protocol and on-chip NICs, this disparity is significant. Technically, this wall could be removed by redistributing the 28 400G links per accelerator, depending on tensor parallelism requirements beyond 4 nodes.
More from Infra
- RL train-sample systems are frontier AI's hardest infra problem — and a trade secret — brianryhuang · 2026-10-16
- Local tuning of GLM-5.3-Flash on 2×GB10 lifts decode from 49 to 58 tok/s — EAccelerate_42 · 2026-10-16
- 3x DGX Spark cluster runs GLM-5.3-Flash locally behind a proxy, 1.75M tokens/day at 119W — Teknium · 2026-10-16
- Optuna creator Akiba: when 2 hyperparameters cost $20M, what's left for BO/HPO — Stefania_druga · 2026-10-16
- Amazon Proposes Per-Request Validated Evidence Caching for Production RAG — _reachsumit · 2026-10-16
- LinkedIn's TransX report: sequence transduction for recsys, +6.02% CTR in A/B tests — _reachsumit · 2026-10-16