Huawei says Ascend has passed Nvidia in China share, deploys 256K-GPU super node
pstAsiatech · x · 2026-10-13
At Huawei Connect, rotating chairman Xu Zhijun and HiSilicon chief scientist Liao Heng briefed media on Huawei's new compute stack:
- Market share: Xu said Ascend's share in China has surpassed Nvidia based on available data; overseas expansion is paused as domestic demand is unmet.
- Peerium architecture: a new compute architecture using nested parallelism (Nested BSP), unified memory addressing, and peer-to-peer interconnect, aiming to make a million processors act as one computer, enabled by the open UnifiedBus high-speed interconnect.
- Atlas 950 SuperPoD: 256TB unified memory, 1024-card scale, 3μs RTT; a 256,000-card cluster is in deployment/testing. The 950PR inference chip is shipping; the 950DT training variant hits volume supply late this year or early next, with many model trainings expected on it from 2026.
- Sizing rationale: models are already at 5T parameters; 6-7 frontier labs in China are expected within two years targeting 10-40T parameter models, making 200K cards a conservative figure given grid power constraints.
More from Infra
- Meta and BlackRock to build 1GW El Paso data center for ~$14 billion — kimmonismus · 2026-10-13
- SubQ claims 128x compute savings at 2M tokens with Subquadratic Sparse Attention — ryanallred · 2026-10-13
- OpenCode 2.0 now runs in Cloudflare Durable Objects via Agents SDK harness — ritakozlov · 2026-10-13
- Cloudflare extends Workers update grace period from 30s to 20 minutes for long-running AI inference — irvinebroque · 2026-10-13
- Musk: China's manufacturing scale beats rest of world combined, but HBM will be harder — elonmusk · 2026-10-13
- ReplaySSM ships in GLM-5.3-Flash and Kimi-K3, speeding decoding up to 2.1x — tri_dao · 2026-10-13