Huawei says Ascend 960 supernode will use 5,500 NPO engines to replace 48,000 800G optical modules
pstAsiatech · x · 2026-09-21
At Huawei Connect, Huawei announced its Ascend 960 supernode will use 5,500 in-house near-packaged optics (NPO) engines in place of 48,000 800G optical modules, entering volume production — though no shipments or third-party verification exist yet.
Key context:
- A July report noted China's supernode designs ran ahead of their components: NPO was prototype-stage and high-end optics (reliant on imported lasers/DSPs) ranked second only to HBM in supply-chain risk. Huawei says it is shortening that list but said nothing on stage about HBM.
- At WAIC 2026 nearly every major Chinese vendor showed a supernode: Moore Threads' 256-GPU system, Enflame+ZTE zero-cable orthogonal design, Biren's optical interconnect targeting 1,024 accelerators, Kunlunxin/Lenovo designs, and Huawei's Atlas 950 SuperPoD (1,024 Ascend chips, design max 8,192).
- Pengcheng Laboratory and the Global Computing Alliance published the first supernode whitepaper, specifying memory-semantic unified addressing, ultra-low latency and ultra-high bandwidth.
- Huatai Securities projects China's domestic supernode market at Rmb 341.4bn ($50.2bn) by 2028.
Related event: Huawei Unveils Ascend 960 Supernode: 4,096 Cards, 8 EFLOPS(5 posts)→
More from Infra
- Artificial Analysis adds market-based GPU pricing to benchmark: B300 leads throughput per dollar — ArtificialAnlys · 2026-11-02
- Qdrant: collection outgrowing RAM pushes rescoring latency from 4.1ms to 43.4ms — qdrant_engine · 2026-11-02
- Instagram Live migrates from in-house RUSH to the IETF Media over QUIC standard — Meta_Engineers · 2026-11-02
- Marvell-Lumentum merger targets AI connectivity as 2028 hyperscaler capex hits $1.5T — BenBajarin · 2026-11-02
- Ex-Google CISO Phil Venables: 70% of vulnerability effort is deployment, not discovery — philvenables · 2026-11-02
- OpenAI, ADNOC and G42 Announce Energy-AI Partnership at ADIPEC 2026 — ingliguori · 2026-11-02