Huawei Challenges NVIDIA with CUDA Alternative
Huawei is taking on NVIDIA's dominance with its new CANN Next framework for Ascend AI chips. The platform reportedly mimics core CUDA concepts to simplify code migration for developers. A full language model has already been trained end-to-end on the new system, proving its viability.
NVIDIA's CUDA platform has a formidable market position, controlling an estimated 92% of the AI datacenter GPU market. This dominance is built on nearly two decades of development, creating a vast ecosystem of optimized libraries and a developer base of over 5 million. This software "moat" makes it difficult for competitors to gain traction, as most AI frameworks are optimized first for NVIDIA hardware. Huawei's Ascend 910B chip, for which CANN is the software stack, offers performance competitive with NVIDIA's restricted A100 GPUs. Manufactured by China's SMIC using a 7nm process, the 910B delivers up to 320 TFLOPS of computing power. Tests have shown the 910B can achieve 80% of the A100's efficiency in large model training, and in some cases, outperform it by 20%. The development of a domestic AI hardware and software stack is a strategic priority for China, amplified by stringent U.S. export controls. These sanctions prevent Chinese firms from accessing top-tier NVIDIA chips like the H100, creating a protected market for Huawei. This has pushed major Chinese tech companies like Baidu and Tencent to purchase the Ascend 910B. To accelerate adoption and build a competitive ecosystem, Huawei has decided to make its CANN (Compute Architecture for Neural Networks) toolkit open source. This strategy aims to encourage developer innovation and reduce reliance on Western technology. The move allows for cross-platform porting and forks, a direct contrast to NVIDIA's proprietary, closed-source approach with CUDA. While CANN supports major frameworks like PyTorch and TensorFlow, building an ecosystem to rival CUDA's maturity and extensive libraries is a long-term challenge. Beyond Huawei, other open-source alternatives are also vying for developer attention, including AMD's ROCm, Intel's oneAPI, and the cross-platform standard OpenCL. The U.S. government has intensified its restrictions, warning that any use of Huawei's Ascend chips, which it claims were likely developed using American technology, could violate export controls and result in severe penalties. This geopolitical pressure aims to slow China's AI progress and creates a stark choice for companies between U.S. and Chinese technology stacks.
