DeepSeek Adds Native Huawei Ascend Support to TileLang and Five Core Libraries

DeepSeek Adds Native Huawei Ascend Support to TileLang and Five Core Libraries

DeepSeek has released native Huawei Ascend support for TileLang, alongside Ascend versions of DeepGEMM, DeepEP, FlashMLA, TileKernels and DeepSelect, according to Leiphone. The package covers kernel development and core computing and communication functions used in large-model workloads. Built on Apache TVM, TileLang lets developers express tiled computations while its compiler handles lower-level operations. The report identifies Ascend 950 as a supported target alongside Nvidia CUDA and AMD ROCm.

For foundation-model labs, this addresses a practical constraint on switching accelerator suppliers: hardware availability alone cannot replace optimized kernels and communication software. The release extends DeepSeek's reported move toward Huawei chips by supplying tools needed to adapt workloads. Its strategic significance lies in reducing dependence on Nvidia-specific software while preserving control over performance-critical operations. However, the Nvidia H100 benchmarks cited in the report do not establish equivalent performance on Ascend or prove that a complete training workload can migrate successfully.

Builders evaluating the toolkit should test their actual matrix multiplication, attention and communication workloads on Ascend before committing to a broader migration. The report identifies remaining challenges in replacing CUDA-dependent training code, adapting collective communications and securing chip deliveries. For investors, the concrete question is whether the software release translates into sustained hardware utilization and reliable operation at scale. Native backend support makes that outcome more plausible; the article provides no end-to-end Ascend training results demonstrating it.

#DeepSeek #HuaweiAscend #TileLang #AIInfrastructure #AICompute