First commercial 1024-card Huawei Ascend 950 SuperPoD delivered to StepFun
teortaxesTex · x · 2026-10-21
On October 15, StepFun reportedly took delivery of the first commercially sold 1024-card Huawei Ascend 950 SuperPoD liquid-cooled supernode, with day-0 adaptation for its Step model.
Key specs:
- Built on the Lingqu high-speed interconnect; single cluster of 1024 Ascend NPUs, fully liquid-cooled cabinets
- 1 EFLOPS FP8 / 2 EFLOPS FP4 compute, designed for trillion-parameter training and high-concurrency inference
- 256TB global unified memory addressing, TB-level card interconnect bandwidth, as low as 3μs end-to-end latency
- Direct card-to-card communication without switches; near-linear scaling
Commenters note Ascend 950 doesn't yet beat Blackwell on economics, but it marks the first real-scale deployment of the supernode product.
More from Infra
- DeepInfer repo needs 64 SM100 GPUs for decode, still lacks a full serving stack — AIFlow_ML · 2026-10-21
- Morgan Stanley: hyperscaler AI capex to hit $3.4T across 2027-2028 with leases included — FinanceYF5 · 2026-10-21
- MiniMax's M3.1 Flash talk: a faster model isn't faster if the job isn't — AI Engineer · 2026-10-21
- Training Factory open-sources QLoRA fine-tuning on 4-bit CQ-quantized models — textclf · 2026-10-21
- DeepSeek open-sources DeepInfer: 155T tokens/day, $38/GPU-hour, >80% margin — Latent Space · 2026-10-21
- Cerebras WSE-3 Turbo packs 4T transistors and 250 PFLOPS into a wafer-scale chip — GregoryDiamos · 2026-10-21