UB-Mesh routes NPU ports across rack, pod and superpod links
bookwormengr · x · 2026-07-22
This follow-up adds more of the paper’s architecture details for Huawei’s UB-Mesh training network.
- Ports are split across different mesh dimensions, so a chip can use links for intra-rack, inter-rack, and pod-level connectivity at the same time.
- The author contrasts this with Nvidia’s more fixed separation between NVLink and PCIe-based connectivity.
- The point is that a flat topology like CloudMatrix 384 or an Ascend 950 superpod needs many more switch chips and optical connectors, while UB-Mesh’s flexible port allocation can reduce that overhead.
More from Infra
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11