Nvidia recently warned developers against using its CUDA on non-Nvidia GPUs, a move where the company tries to fend off other competitors. It also highlights another challenge in China's efforts to become self-sufficient in AI chip development.
According to Tom's Hardware, Nvidia updated its end-user license agreement for CUDA 11.6, warning against developers using CUDA on third-party GPUs through "translation layers," which converts one set of codes into another, on third-party GPUs. The agreement stated that reverse engineering, decompiling, or dissembling any results generated using this SDK and translating them onto non-Nvidia platforms are disallowed.
Tom's Hardware reported that Nvidia's move may target the open-sourced ZLUDA or try to stop China-based GPU developers from leveraging CUDA through the translation layers.
South China Morning Post quoted Su Lian Jye, chief analyst at Omdia, saying CUDA plays a crucial role in Nvidia's ecosystem in locking in its customer base, adding if the compatibility with CUDA was undermined, the costs of integrating codes would be so much higher.
Chinese media Observer News quoted Moore Threads' statement saying the company's MUSA/MUSIFY is independent of CUDA and does not involve Nvidia's terms regarding the end-user license agreement, reassuring developers that they can use it with confidence. China-based GPU provider Biren Technology told Observer News that its BIRENSUPA was independently developed and had nothing to do with Nvidia's updated license agreement.
In the meantime, Huawei, viewed by Nvidia as one of the primary competitors in AI chips, offered the Compute Architecture for Neural Networks (CANN), which is still in its early stages of development and lags behind CUDA in user adoption.
Multiple Chinese media reported that Nvidia may not aim at China-based competitors but at ZLUDA, an open-sourced platform. An unnamed source told Chinese media Jiamian that CUDA is a closed-source platform and cannot be used on non-Nvidia products. The updated terms did not change anything. The source added that Nvidia might attempt to confirm the proprieties of CUDA when ZLUDA became open-sourced after AMD withdrew funding. AMD and Intel provide ROCm and OneAPI, respectively.
