DeepSeek open-sources full Ascend software stack: TileLang, DeepGEMM, DeepEP and more
DeepSeek · wechat · 2026-09-30
DeepSeek officially open-sourced its infrastructure stack for Huawei Ascend, mirroring its NVIDIA-platform components one-to-one: the TileLang compiler (a simpler, higher-level alternative to CUDA that already carries most operator implementations in DeepSeek V4 training), DeepGEMM-Ascend for matrix ops, DeepEP-Ascend for large-scale cross-device communication, TileKernels, FlashMLA for sparse attention, and DeepSelect.
In key benchmarks the components reach near-hardware-limit compute and communication performance. Huawei worked closely with DeepSeek, jointly optimizing a 128-card supernode方案 based on Ascend 950. DeepSeek hopes the move sets an example for building usable software ecosystems on more AI chips.
Related event: DeepSeek Open-Sources Full Ascend AI Infrastructure Stack(5 posts)→
More from Infra
- Quantized softmax attention pretraining: only +0.004 nats loss gap at K=16 with the right calibration — illinois · 2026-09-30
- xLLM training infra open-sourced with xattn attention backend and xBridges toolkit — HongyiWang10 · 2026-09-30
- Auto-research loop on 120 B300s finds 40% Kimi K3 inference gain for $9,176 — bookwormengr · 2026-09-30
- Cerebras to bring 'world's fastest inference' to General Compute — beffjezos · 2026-09-30
- 8% of Asia-to-US air freight is now data center parts — 30 full freighters a day — yacineMTB · 2026-09-30
- mradermacher quants get Gemma 26B to 75 tok/s on 2x RTX 4060 8GB — Spiritual_Impress_30 · 2026-09-30