Huawei: 1000+ Ascend 910C supernodes deployed, 40+ LLMs natively pretrained

teortaxesTex · x · 2026-10-04

At HUAWEI CONNECT 2026, Ascend Computing president Zhang Dixuan detailed Huawei's AI compute progress:

• Over 1000 Ascend 910C supernode deployments with 650M+ online card-hours; 40+ LLMs natively pretrained on Ascend, claimed as China's only pretraining-capable domestic compute platform.
• Ascend became an officially supported PyTorch backend in July, the sole Chinese hardware vendor.
• Open-sourced CANN now spans 30+ SIGs and 70+ projects across 90+ communities, with 5000+ monthly active developers.
• The UnifiedBus (LingQu) interconnect and "supernode + cluster" architecture target bandwidth, latency, and unified memory addressing to serve scaling from 1T to 10T parameters and agentic/embodied AI workloads.

Original post →

More from Infra

• Musk: xAI to hit 10GW of compute by end of next year, sees AI inference moving to space — beffjezos · 2026-10-04
• BF16 rounding breaks a conservation law, blowing up FlashAttention gradients late in training — HongyiWang10 · 2026-10-04
• Getting PyTorch CUDA training running on BC-250 boards, captured as an image — redfoxkiller · 2026-10-04
• Huawei 950 super-node claims seamless scaling from 550B/1.6T up to 10T-class models — teortaxesTex · 2026-10-04
• Huawei's Ascend 950DT TDP hits 950W per card; B200 ~4.4x denser in FP8/W — teortaxesTex · 2026-10-04
• The curse of 64GB RAM: Strata pushes local Qwen3.8-Flash-Next to 60 t/s but hogs system memory — Cautious_Chicken_604 · 2026-10-04

Infra channel →