Inside Huawei's Ascend 960 Supernode: 4,096 Cards, Unified Memory, 2.75x MFU
智东西 · wechat · 2026-09-21
智东西这篇文章拆解了华为在 HC2026 上发布的昇腾960超节点。背景是：10万卡集群中卡间通信可能消耗40%以上训练时间、MFU往往不到30%，超节点通过统一内存编址把数千卡组织成逻辑上的「一台计算机」，华为仿真显示4000卡超节点的 MFU 可达传统8卡集群的2.75倍。
Related event: Huawei Unveils Ascend 960 Supernode: 4,096 Cards, 8 EFLOPS(5 posts)→
More from Embodied
- Robot only dances to songs it likes: FFT-based taste check plus BPM prediction — wavefnx · 2026-11-02
- Musk spotlights ALS patient controlling wheelchair with his mind via Neuralink — elonmusk · 2026-11-02
- Fetchbot, a homemade robot, takes its first drive — _Stocko_ · 2026-11-02
- Day 8 of PhoneGirl humanoid training: it walks, but hilariously — _Stocko_ · 2026-11-02
- Local AI computer maker Autonomous ships to 5 countries, prices up to $99,700 — dee_hw · 2026-11-02
- Drone-captured 400-ft guyed tower rebuilt as 3DGS via Metashape and LichtFeld Studio — janusch_patas · 2026-11-02