56.8 ÷ 6.7 ≈ 8.5：华为昇腾超节点为什么要把 160 个机柜连成一台计算机

系列第六期：华为昇腾超节点——单颗芯片不够强，就用数量来补；数量带来距离、可靠性、能耗三个新问题，每一代超节点都在回答上一代的问题。 Episode 6, Chinese narration with Chinese captions burned in. Same ~8 EFLOPS FP8: about 7 NVIDIA Vera Rubin racks, or 160 Huawei Atlas 950 racks — but the 160 racks are one scale-up computer. We walk through CloudMatrix 384 (16 racks, 5,000+ optical modules; wins on scale, loses on efficiency), Atlas 950 (8,192 chips; 56.8× chips for 6.7× compute ≈ 1/8 per die; 15× memory; 950PR vs 950DT memory split), reliability (UnifiedBus, ~100 ns optical failover, the copper-first UB-Mesh design), power (Atlas 960E: half the chips; 48,000 pluggables → 5,500 Hi-ONE NPO engines, ›550 kW saved; why NPO not CPO), and why the two roads — set by each side's scarcest resource — are converging. Chapters 00:00 7 racks vs 160 racks 00:27 Seven computers vs one 01:41 1 · Not strong enough? Use more 02:44 2 · Problem one: distance (CloudMatrix 384) 03:47 Wins scale, loses efficiency 04:40 3 · Twenty times bigger: Atlas 950 05:17 56.8 ÷ 6.7 06:19 The real advantage: memory 07:34 950PR vs 950DT 08:42 4 · Problem two: reliability 09:53 UnifiedBus 10:49 ~100 ns failover vs Google's mirrors 11:40 UB-Mesh: copper first 13:10 5 · Problem three: power (Atlas 960E) 14:22 Hi-ONE: 48,000 → 5,500 15:32 Why near-packaged, not co-packaged 16:58 6 · Two roads, converging 17:03 Scarcity decides the road 18:30 Why the single computer keeps growing (MoE) 19:50 Converging 20:47 What to watch 22:23 The network between the chips Sources: Huawei Connect 2025/2026 keynotes and press releases · SemiAnalysis (CloudMatrix 384) · UB-Mesh (arXiv 2503.20377) · Serving LLMs on CloudMatrix384 (arXiv 2506.12708) · DeepSeek-V3 report · NVIDIA Vera Rubin POD blog · The Register · TechNode · Bloomberg · Reuters. Huawei's comparison counts NVIDIA's rack as 144 GPU dies (NVIDIA now names it NVL72, by package). UB-Mesh is a research design. Facts and engineering trade-offs only; not investment advice. #华为 #昇腾 #超节点 #AIInfrastructure #ScaleUp #DataCenter