DeepSeek Open-Sources TileLang, a CUDA Rival for Huawei Chips
By @sharedot · · 8 pages
- Programming
- Deepseek
- Huawei Ascend
- Tilelang
- Open Source
DeepSeek released six open-source tools, led by TileLang, that let developers program Huawei Ascend AI chips as a direct challenge to Nvidia's CUDA ecosystem.
What shipped
On September 30, DeepSeek announced via its official WeChat account that it is open-sourcing a full programming toolkit for Huawei's Ascend AI accelerators, built jointly with Huawei.
The release centers on TileLang, a high-level language for writing the performance-critical kernels that determine model speed, and includes DeepGEMM for matrix operations, DeepEP for cross-chip communication, TileKernels for vector computation and memory access, FlashMLA for long-context attention, and DeepSelect for data filtering.
DeepSeek says each Ascend component corresponds one-to-one with tools it previously built for Nvidia GPUs, so developers can migrate workloads with far less duplicated effort. Reuters independently reported the partnership.
Why TileLang matters
TileLang was originally developed by researchers at Peking University, and DeepSeek reports it has used the language for about a year, first testing it on older Nvidia chips. According to The Decoder, TileLang is now DeepSeek's main tool for its AGI work.
DeepSeek argues that any independent, self-controlled GPU software ecosystem must start with a universal high-level language that is easy to program yet still reaches the hardware's full performance, and it describes TileLang as offering "a simpler programming model" than CUDA.
The Ascend implementation wraps Huawei's Ascend C instructions in that higher-level model, lowering the barrier that has historically made leaving Nvidia hardware expensive.
The evidence in the repo
According to Dataquest, the public TileLang-Ascend project publishes its own performance measurements: average performance around 0.98 times hand-written Ascend C for typical GEMM workloads, roughly 0.96 times for vector operators, and about 0.95 times for cube-vector workloads such as Flash Attention.
Those are project-published benchmarks rather than independent tests, but they show working code, not just an announcement. The repository also records DeepSeek V4 kernels for Ascend released in April 2026, indicating the software effort predates the announcement.
The project documents an Ascend-specific NPU intermediate representation connecting to the open-source MLIR compiler ecosystem.
Why this is the surprising move
The strategic significance goes beyond one release. Nvidia's dominance rests not just on silicon but on a software ecosystem — Nvidia says more than 7.5 million developers use CUDA and its other tools — that has made rival chips hard to adopt even when hardware looked competitive.
Startup Fortune reports that DeepSeek gave early access to its V4 model to Chinese chipmakers rather than Nvidia or AMD, and that ByteDance, Tencent, and Alibaba then rushed to order Ascend 950 processors.
Open-sourcing the toolkit is the second half of that play: first prove a frontier model runs on domestic silicon, then hand every other Chinese lab the software to follow.
The stakes: supernodes and export controls
DeepSeek and Huawei also co-optimized a supernode system of 128 Ascend 950 chips, tuning computation and chip-to-chip communication, with DeepSeek claiming performance in some test cases has approached hardware limits.
The work is unfolding under US export controls that restrict Chinese access to Nvidia's most advanced processors; Huawei's rotating chairman Eric Xu said the company cannot accept a future that hinges on whether others are willing to sell chips to China.
BigGo Finance reports DeepSeek is building an Inner Mongolia data center with more than 160,000 Ascend processors and is finalizing a capital raise near $74 billion ahead of a potential STAR Market IPO.
What comes next
The gap remains substantial: analyst commentary cited by BigGo Finance warned that shifting workloads away from CUDA would likely take years, and SemiAnalysis still found Nvidia well ahead on multistep agent workloads as recently as August.
Nvidia shares dipped less than 1% on the announcement, and customers are still signing multiyear contracts for older H100 chips, so this is not an overnight displacement story.
The signal to watch, as Startup Fortune puts it, is adoption outside Huawei and DeepSeek — whether smaller Chinese labs and startups start building on TileLang instead of CUDA. Huawei, meanwhile, has said its next-generation Ascend 960DT ships in Q1 2027 with a roadmap extending to 2029.
Sources
- the-decoder.com › China's AI industry closes ranks as Deepseek ships open-source software for Huawei's Ascend chips
- finance.biggo.com › DeepSeek and Huawei Unveil Open-Source Toolkit to Loosen Nvidia's CUDA Grip
- dqindia.com › DeepSeek expands Huawei Ascend push with six open-source AI tools
- startupfortune.com › DeepSeek open sources chip tools that could let Huawei replace Nvidia in China