DeepSeek Open-Sources Full Ascend Toolchain, Taking Aim at CUDA
On September 30, DeepSeek officially open-sourced six infrastructure projects for Huawei Ascend, mirroring its previously open-sourced NVIDIA components one-to-one and completing a full port from NVIDIA GPUs to Ascend chips. Per @rohanpaulai, matrix kernels reach 99.8% of the chip's theoretical peak; @lxfater also reports several components near hardware limits. Bloomberg, Reuters and SCMP uniformly frame the move as a challenge to NVIDIA's CUDA moat.
Confirmed
- DeepSeek's official post confirms six projects: TileLang for Ascend (a high-level compiler wrapping AscendC, claimed easier than CUDA while exploiting chip performance), DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA and others; SCMP counts six software modules in total.
- @rohanpaulai adds that TileLang is DeepSeek's in-house kernel DSL powering most operators in V4 training, with the TileKernels repo offering dozens of deeply optimized operators covering MoE routing, quantization and other LLM workloads.
- Per @aigclink, this stack was built for training the V4 model family and is now fully open-sourced.
- Per @kimmonismus (Bloomberg) and @markk (Reuters), DeepSeek cooperated with Huawei on Ascend programming tools; TileLang supports Ascend 950-specific optimization with the compiler handling scheduling and synchronization, positioned as a potential CUDA alternative. The TileLang repo has 7.7k GitHub stars.
- Per @pstAsiatech (SCMP), the goal is a self-contained GPU software ecosystem independent of CUDA, reducing reliance on NVIDIA under US export controls; @KeyNaive2982 notes the tools aim to help Huawei chips replace NVIDIA GPUs in training and inference.
- @teortaxesTex argues that with DeepSeek's kernel expertise, this makes the Ascend training ecosystem fully usable; @vista8 calls it a key signal of DeepSeek's infrastructure merging with the domestic compute ecosystem.
Why it matters
- DeepSeek's prior open infrastructure targeted NVIDIA GPUs; this full Ascend port, released officially with matrix operators at 99.8% of peak, could sharply lower the barrier to reproducing large-model training on Ascend hardware. Bloomberg and Reuters position TileLang as a CUDA-alternative signal, while SCMP stresses building an independent software ecosystem under export controls.
2026-09-30 ~ 2026-10-01 · 17 related posts
Primary sources
- [source] DeepSeek open-sources full Ascend software stack: TileLang, DeepGEMM, DeepEP and more — DeepSeek · 2026-09-30
- DeepSeek open-sources foundational components for Huawei Ascend AI stack — vista8 · 2026-09-30
- [source] DeepSeek open-sources full Ascend AI training stack, mirroring its NVIDIA components — aigclink · 2026-09-30
- Full repo list: DeepSeek's Ascend stack — TileLang, DeepGEMM, DeepEP, FlashMLA — aigclink · 2026-09-30
- DeepSeek Ships TileLang Ascend Support, Signaling Ascend Now Viable for Training — teortaxesTex · 2026-09-30
- DeepSeek open-sources Huawei Ascend toolkit with TileLang support, challenging Nvidia's CUDA — kimmonismus · 2026-09-30
- DeepSeek and Huawei team up to chip away at Nvidia's CUDA moat with TileLang open source — mark_k · 2026-09-30
- DeepSeek open-sources six tools to help Huawei Ascend chips replace Nvidia in AI — pstAsiatech · 2026-09-30
- DeepSeek ships six tools to help Huawei Ascend replace Nvidia; TileLang adds Ascend 950 support — pstAsiatech · 2026-09-30
- DeepSeek Open-Sources Six Tools to Help Huawei Ascend Chips Replace Nvidia — pstAsiatech · 2026-09-30
- DeepSeek reportedly building software for Huawei Ascend 950 chips, open-sourcing six CUDA-challenging tools — IntCyberDigest · 2026-09-30
- DeepSeek open-sources tools to help Huawei chips supplant Nvidia in AI, SCMP reports — KeyNaive2982 · 2026-10-01
- [source] DeepSeek open-sources six Huawei Ascend projects, kernels hit 99.8% of chip peak — rohanpaul_ai · 2026-10-01
- DeepSeek releases TileKernels: dozens of LLM kernels near hardware limits, now on Ascend — rohanpaul_ai · 2026-10-01