DeepSeek has published two new open-source repositories targeting Huawei’s Ascend accelerators, for matrix multiplication and mixture-of-experts communication. Their familiar API shapes could make experimenting beyond Nvidia’s CUDA ecosystem less of a rewrite, although the supported hardware is limited.
DeepSeek Watch analysis
What happened
The two new projects are DeepGEMM-Ascend and DeepEP-Ascend. TechRadar reports that the announcement also included Ascend-targeted updates to TileKernels, DeepSelect and FlashMLA, which are existing projects rather than new repositories.
The report says DeepGEMM-Ascend requires Ascend 950-series hardware and CANN 9.20. DeepEP-Ascend requires Ascend 950 silicon with UBMEM connectivity. It also says DeepGEMM-Ascend keeps the Python package name and API shape of its CUDA counterpart, while DeepEP-Ascend aligns its public buffer interfaces with the Nvidia version.
Why it matters
That compatibility work is aimed at a stubborn problem in AI infrastructure: teams have invested years in CUDA code, tools and expertise. Making some interfaces familiar will not make a different accelerator effortless to adopt, but it could reduce the amount of code that needs rewriting before developers can test one.
The catch is practical and considerable. The reported hardware and toolkit requirements limit who can run these projects now, and matching interfaces do not establish matching performance or a straightforward production migration. Open-source code is a useful invitation; it is not a delivery service for scarce silicon.
Our read
The most interesting part is not a claim that Ascend has already displaced Nvidia. It is DeepSeek contributing working software that may lower the switching cost for developers willing and able to access the hardware. That is a concrete step towards a more usable alternative, with plenty still to prove in broader deployments.
For teams exploring non-CUDA options, the useful next step is to check the exact chip, CANN version and connectivity requirements before treating these repositories as ready-to-run alternatives. The real test is whether outside developers can use them, reproduce results and maintain them beyond the announcement.
What to watch
- Whether DeepSeek or Huawei broadens hardware and toolkit support.
- Whether developers outside the two companies report successful builds and deployments.
- How the Ascend projects perform on comparable workloads, and how much porting work remains.
Discussion spark: If a new accelerator can run familiar interfaces but still requires scarce hardware and a specific software stack, has it meaningfully lowered switching costs or mostly moved the barrier elsewhere?
Sources and evidence
not affiliated with or endorsed by DeepSeek