Huawei's UnifiedBus lifts 100k-card MFU from 20% to 35%, unveils Ascend 960 supernode
智东西 · wechat · 2026-09-19
At Huawei Connect 2026, Huawei launched the Ascend 960 chip, Ascend 960 supernode, and OceanStor M900 storage, all built around its UnifiedBus (LingQu) interconnect protocol.
Traditional interconnects leave 100k-card clusters below 20% MFU. UnifiedBus unifies CPU, NPU, DPU, DDR, and SSD under one protocol with global memory addressing, boosting interconnect bandwidth from 100GB to TB-class, cutting RTT latency from 7μs to 2μs (70% lower), and raising 100k-card MFU to 35%. With optical-electrical interconnect, a single cluster can scale to 1 million AI processors.
The Ascend 960 supernode reaches 4,096 cards with 8 EFLOPS FP8 and 1PB HBM; air- and liquid-cooled versions ship in 2027 Q2/Q3. The Atlas 650E all-in-one connects two machines over UnifiedBus for 16 NPUs running trillion-parameter models locally, with 40%+ throughput gains. On the ecosystem side, the open-sourced CANN now has 5,200+ monthly active developers and 26 industry-specific operator libraries.
More from Companies & People
- Dev spent 24 hours ranking companies by Jev's transformational potential and pitching the top ones — vaibhavbetter · 2026-09-19
- Harvard pushes AI into classrooms with Claude access as faculty backlash grows — luislamb · 2026-09-19
- Blogger: most engineers are just vibe coders, real skills lie in fundamentals — ashishllm · 2026-09-19
- AI Twitter's Overton Window Shifted in 30 Days: METR Buzz, GPT-6 Demos and Fly Memes — dejavucoder · 2026-09-19
- Anthropic's Accenture embedded evaluator announcement roasted for its medieval-cat artwork — dejavucoder · 2026-09-19
- OpenAI pulled 25% of production engineers to let Astra hunt its own vulnerabilities to saturation — kevinnbass · 2026-09-19