NVIDIA Prohibits the Use of CUDA on Third-Party Accelerators

NVIDIA has issued a warning to developers to avoid the utilization of CUDA on graphics processors from third-party manufacturers, possibly in a move to fend against competition. This highlights the difficulties faced by China in developing its own chips for artificial intelligence (AI) systems. However, the real reason may not be what first meets the eye.

NVIDIA’s Update on Licensing Agreement with CUDA 11.6

Following the release of CUDA 11.6, NVIDIA updated its license agreement warning developers against applying this technology on graphics processors from third-party developers using “translation layers” capable of converting one code into another. The manufacturer emphasized that the reverse engineering, decompilation, or disassembly of its SDK for translation onto platforms other than NVIDIA is prohibited.

Chinese GPU manufacturers’ Response to NVIDIA’s policy

Chinese GPU manufacturer, Moore Threads, has assured that its solutions, MUSA/MUSIFY, are not linked with CUDA, and thus, NVIDIA’s updated policy does not impact its users. Similarly, Biren Technology, clarified that its technology, BIRENSUPA, was independently developed and therefore untouched by NVIDIA’s actions. NVIDIA’s named competitor in the AI chip sphere, Huawei, offers its own neural network architecture, CANN. However, it is still in the early stages of development and lags behind CUDA in user adoption.

Is NVIDIA’s Policy Update Aimed at Open Platform ZLUDA?

Chinese media speculates that NVIDIA’s updated policy may actually be aimed not at its competition but at open platform ZLUDA. Being a closed-source platform, CUDA technically could never be used on third-party equipment, and NVIDIA is just reinforcing this stance. This move comes after AMD pulled funding from the ZLUDA project, leading the developer to make it open-source. Concurrently, AMD and Intel provide platforms ROCm and OneAPI respectively.

Related Posts