DeepSeek's Ascend ports swap a CUDA dependency for Huawei's CANN

Both ports split the stack at the same line. The Python functions a developer calls stay fixed, and the kernel code under them is written separately for each vendor ^[7][17]. Kernels are the small, optimized programs that do most of the repetitive math in training and inference ^[13]. That is where hardware-specific work piles up. For code already written against DeepGEMM or TileKernels, moving to Ascend becomes a backend choice and the call sites stay as they are ^[17].

I think DeepSeek picked the right layer. Nvidia's lead rests partly on roughly two decades of developers programming its GPUs through CUDA ^[5]. The Neuron describes the switching cost plainly: every incompatible piece of a stack "creates another engineering bill" ^[8].

The vendor dependency moves down one layer. DeepGEMM-Ascend requires Ascend hardware and CANN, Huawei's Compute Architecture for Neural Networks toolkit ^[9]. TileKernels lists CUDA for its Nvidia backend and CANN for its Ascend backend ^[18]. A lab that leaves Nvidia through these libraries trades one vendor toolkit for another. Open source, in this release, means anyone can read the code that calls into Huawei's toolkit ^[9]. The Neuron's report notes that open source does not automatically mean freedom from vendor dependence ^[16].

The Neuron wrote that "portability at the API layer is different from parity underneath it" ^[12]. A matched API tells a lab its calls will run on Ascend. Speed at full model size is a separate measurement, and so is whether a long job stays up. The report argues that kernel benchmarks are only part of the test ^[14]. I'd expect any published kernel figure to carry over only to labs running similar matrix shapes on similar Ascend hardware and CANN versions.

According to The Neuron, independent developers have not yet measured how much engineering work the tools remove from choosing Ascend ^[10]. The same report says whether Ascend becomes a dependable alternative depends on chip supply, engineering effort, reliability and full-scale model performance ^[11]. Supply sits outside the code. The report argues that better software cannot solve a chip shortage ^[15].