Huawei Unveils Ascend 960 Chips and SuperPoD Stack Aiming at Million-Card Clusters
智东西 · wechat · 2026-09-17
At Huawei Connect 2026, Huawei unveiled a full AI infrastructure stack built on its UnifiedBus interconnect:
- Ascend 960DT ahead of schedule: 2 PFLOPS FP8 / 4 PFLOPS FP4, double last-gen performance, 288GB HBM at 9.6TB/s, ready Q1 2027; 970 and 980 to follow annually in 2028–29.
- Ascend 960 SuperPoD: first NPO-based supernode — 4,096 cards, 8 EFLOPS FP8, 1PB HBM, 99.8% availability — paired with the world's first mass-produced 7.2Tbps NPO optical engine (Hi-ONE).
- Kunpeng SuperPoD: 4,096 nodes, 256TB unified memory pool, 30× faster agent sandbox startup, 300K TPS vector search.
- OceanStor M900: KV Cache in one hop, >50% lower latency, 16× SSD lifespan.
- Networking: UnifiedBus + two-layer CLOS supports up to 512K Ascend 960 cards and million-processor clusters; 10K-card cluster MFU raised from 20% to 35%.
CANN now has 5,000+ monthly active developers (61% external); Kunpeng ecosystem exceeds 4.16M developers. Huawei also outlined on-device AI plans: 30B Pangu MoE shipping on Mate X2,端侧 models potentially reaching 70–100B by 2030.
More from Embodied
- Sofía Dudas Kicks Off SSAD 2026 Day 3 With Talk on World Models for Autonomous Driving — abursuc · 2026-09-17
- Ropedia launches academic partner program offering HOMIE Gen2 kits for physical AI data — MengdiWang10 · 2026-09-17
- Doubao launches in-car cockpit assistant with multimodal vehicle control — xiaohu · 2026-09-17
- World Synesthesia Model Enables 1-Minute Noise-Robust In-Hand Rotation, CoRL 2026 Paper — 机器之心 · 2026-09-17
- ActionPiece rethinks action tokenization for VLA models, hits 94.8% on LIBERO — DeepCybo · 2026-09-17
- Tesla Optimus passes CAPTCHA designed to prove you're not a robot — TansuYegen · 2026-09-17