Huawei details training DeepSeek V4 on Ascend 950 supernodes: 273K tok/s on 2 racks
teortaxesTex · x · 2026-10-17
Huawei keeps using DeepSeek as the default reference for its Ascend stack. Per the cited CANN reference-practice slide:
- DeepSeek V4.1 Flash (550B): trains on 1 supernode (128 NPUs), all EP communication within the node, using FSDP128/EP64-style parallelism.
- V4 Pro (1.6T): 2 supernodes (256 NPUs) with FSDP2 across nodes, no pipeline bubbles.
- 10T-scale models: 8 UBL128 units, 1,024 NPUs with PP, keeping EP communication inside each UBL to cut cross-node overhead.
Performance highlight: DeepSeek V4 Flash hits 2,131 tokens/s/NPU training throughput on a PC16 cluster (64 NPUs, 4K seq len). teortaxesTex notes V4.1 Flash does 273K tok/s on just 2 racks (under 24B tokens/day), but the 4K sequence length differs from DeepSeek's own setup, making it only good for finetuning.
More from Infra
- NVIDIA releases multimodal reranker on Hugging Face for multilingual vision RAG — solyarisoftware · 2026-10-17
- Databricks research team now runs experiments with an agent swarm, training volume up 8x in a year — matei_zaharia · 2026-10-17
- e/acc figure backs local AI as 'civilizational infrastructure' against cloud gatekeeping — beffjezos · 2026-10-17
- GPU prices rose with age: RTX 50 series up 66% YoY, RTX 5090 up 132% on memory shortage — jasonkneen · 2026-10-17
- Onchain AI inference surging: Venice hits 500B tokens/day as Bittensor expands subnets — 0xJeff · 2026-10-17
- Anthropic reportedly offers Blackstone 10%+ returns to fund AI chip deal — Polymarket · 2026-10-17