Nvidia Dominates China's Model Development Scene Despite Huawei's Chip Push

China's semiconductor self-sufficiency drive has been ongoing for some time now, but despite Huawei's efforts to develop its own chips, Nvidia remains a dominant player in the country's model development scene. The main reason for this is not just that Chinese-made chips can't handle the workload, but also that switching hardware would require significant changes to the software systems built around Nvidia's CUDA platform.

This creates an engineering challenge for model developers who are accustomed to using CUDA and may need to rewrite substantial portions of their existing code before moving their training workloads to Huawei's Ascend chips. For example, James Wang, a researcher at a Shanghai-based university, estimated that switching to Ascend could increase the time and cost of the process by at least 50%.

The burden varies depending on how open the model is and how mature the supporting software ecosystem has become. A Beijing-based engineer involved in moving model workloads to Ascend told SCMP that open-source models can be migrated more easily because developers can use existing ecosystem support, but for models whose underlying source code is not publicly available, it could require about 10 engineers and more than six months of additional work to make the switch.