TileLang Opens as Multi-Backend Alternative to CUDA

DeepSeek and Huawei open-sourced TileLang, a portable multi-backend language and libraries challenging Nvidia's CUDA.

@sharedot

TileLang Opens as Multi-Backend Alternative to CUDA

By @sharedot · · 8 pages

  • Programming
  • Open Source
  • Tilelang
  • Ascend
  • Cuda

DeepSeek and Huawei open-sourced TileLang, a portable multi-backend language and libraries challenging Nvidia's CUDA.

What Happened on September 30

DeepSeek and Huawei released a free, open-source programming toolkit for Huawei's Ascend AI chips on September 30, 2026, built around TileLang — a high-level language for writing optimized AI kernels.

The package includes compute and communication libraries plus native Ascend 950 support, with Huawei providing full support during development. The two companies jointly optimized computation and communication on a supernode system built around 128 Ascend 950 chips.

Six Modules, One Portable Stack

According to Tech Times, the package contains six modules now optimized for Ascend 950: DeepGEMM-Ascend for matrix multiplication, DeepEP-Ascend for large-scale cluster communication, TileKernels for vector operations and memory access, FlashMLA for long-context processing, DeepSelect for data filtering, and TileLang itself.

DeepGEMM-Ascend supports BF16, FP8, and FP4 operations and keeps the same programming interfaces as DeepSeek's existing DeepGEMM library, so developers can keep familiar APIs when moving to Ascend.

DeepEP-Ascend routes data to different experts in mixture-of-experts models and combines their outputs.

Why TileLang Matters

TileLang operates at a higher abstraction level than CUDA: rather than threads, the fundamental unit is a tile, and the language decouples scheduling from dataflow so the compiler handles synchronization on the target hardware.

Tech Times reports that TileLang already supports Nvidia CUDA, AMD ROCm, and Apple Metal as backends, with Ascend 950 now added — meaning code written on Nvidia hardware can run on Ascend without rewriting a single kernel.

DeepSeek described the language as offering a simpler programming model than Nvidia's CUDA, aimed at improving development efficiency, and uses it as its core tool for AGI research.

The Attack on a Two-Decade Moat

CUDA has been Nvidia's most durable competitive advantage since its first public release in 2007, with BusinessKorea noting the platform is used by 4 million developers worldwide.

Tech Times reports Nvidia said at GTC 2026 that more than six million developers now build on CUDA, alongside more than 400 GPU-accelerated tools.

The question the release poses, per Tech Times, is whether an open-source, multi-backend abstraction layer can start the same community flywheel that built CUDA's moat.

Gregory Allen of CSIS told Congress in April 2025 that DeepSeek's expertise could improve Huawei's CANN ecosystem, while cautioning that shifting workloads away from CUDA would likely take years.

The Training Gap Remains Open

The software push does not resolve the hardest problem: DeepSeek still uses Nvidia chips for model training while Huawei chips are primarily used for inference, BusinessKorea reports.

Tech Times adds that when DeepSeek attempted to train an earlier model on Huawei silicon, the effort stalled due to persistent technical difficulties, according to the Financial Times.

Huawei expects its AI systems to see widespread use in model training by 2027, a claim the company frames as forward-looking. DeepSeek plans to deploy at least 160,000 Ascend 950DT chips at its data center under construction in Inner Mongolia.

What Comes Next

The release signals continued investment in software alternatives to Nvidia's ecosystem, though CUDA remains the dominant platform globally.

Tech Times notes Nvidia acknowledged in its latest annual filing that competitors have built larger developer ecosystems to challenge it, and CEO Jensen Huang said in a May 2026 CNBC interview that Nvidia has largely conceded the China market to Huawei.

Huawei announced plans for next-generation Ascend chips in 2028 and 2029, with rotating chairman Eric Xu saying it is unacceptable to allow external forces to determine China's technological future. TileLang's multi-platform support positions it as a bridge rather than a replacement.

Sources

  1. news.lavx.hu › DeepSeek and Huawei open-source Ascend programming tools to challenge Nvidia's CUDA dominance
  2. techtimes.com › DeepSeek and Huawei Open-Source Toolkit Lets Developers Ditch CUDA Without Rewriting Code
  3. businesskorea.co.kr › DeepSeek, Huawei Join Forces to Challenge Nvidia's CUDA Dominance

More on Programming