China's Alternative to CUDA? Inside DeepSeek's DeepGEMM-Ascend

Building an AI chip is only half the problem—the software has to know how to feed it numbers. On September 30, 2026, DeepSeek open-sourced DeepGEMM-Ascend, porting its high-performance matrix multiplication kernels from NVIDIA to Huawei Ascend hardware while keeping the top-level Python API unchanged. 🔔 Subscribe: https://www.youtube.com/channel/UC0DZj1PNa_Fp0MU6uPSKv5w?sub_confirmation=1 💙 Become a Member: https://www.youtube.com/channel/UC0DZj1PNa_Fp0MU6uPSKv5w/join 🐦 Twitter/X: https://x.com/cloud_codes 💬 Discord: https://discord.gg/mhgXXmhrh In this technical breakdown, Cloud Codes explains what DeepGEMM-Ascend does and what it does not do: • The Hardware Seating Plan: Why matrix data must be rearranged during data movement into specialized blocks to feed Huawei's Cube computation units. • Pipelining & Memory: How coroutine-based pipelining overlaps data delivery with active computation so execution units never sit idle. • API Continuity: How familiar cuBLASLt calls and matrix functions map directly to Huawei's ACLNN backend without forcing developers to rewrite application code. • Hardware Validation: The release was built and verified specifically for Ascend 950 running CANN 9.20 and torch_npu. • The 99.8% Benchmark: Understanding DeepSeek's reported 99.8 percent hardware limit on BF16 operations (a kernel utilization figure under cold L2 profiling, not an application-wide speedup). • Kernel Library vs. CUDA: Why porting matrix math is a huge milestone for non-NVIDIA silicon, but does not replace the full CUDA compiler, runtime, and developer toolchain. 🔗 Verified Sources: • DeepGEMM Ascend Repository (GitHub): https://github.com/deepseek-ai/DeepGEMM-Ascend • Upstream DeepGEMM (CUDA): https://github.com/deepseek-ai/DeepGEMM • Python API Entry Point & ACLNN Aliases: https://raw.githubusercontent.com/deepseek-ai/DeepGEMM-Ascend/main/deep_gemm/__init__.py • Huawei CANN Data Movement Documentation: https://www.hiascend.com/document/detail/en/CANNCommunityEdition/910/API/ascendcopapi/docs/en/api/SIMD-API/basic_api/data_move_guide/overview/data_movement_concept.md • NVIDIA CUDA Toolkit Architecture: https://developer.nvidia.com/cuda/toolkit #deepseek #huawei #ascend950 #cuda #aihardware #machinelearning #gpu #cloudcodes User Queries: deepseek deepgemm ascend github release is deepseek building a cuda replacement for huawei how does deepgemm ascend work on huawei 950 difference between deepgemm and cublaslt huawei cann 9.20 torch_npu deepseek deepgemm ascend 99.8 percent hardware limit can huawei ascend run deepseek models natively huawei cube unit matrix data layout conversion why is porting cuda kernels to huawei hard china domestic ai chip software ecosystem deepseek