In a recent development, there has been an analysis of the legality surrounding the creation of alternative implementations of CUDA technology. One such implementation is the Zluda 3 package, which enables the launching of CUDA applications on AMD GPUs. This comes in light of Nvidia’s addition to the user agreement (EULA) of CUDA Toolkit 11.5, which now prohibits reverse engineering, decompilation, and disassembly of any output generated by CUDA tools with the aim of broadcasting such artifacts on platforms other than NVIDIA. Notably, providing the source code of CUDA programs, along with tools for translating CUDA programs to the AMD ROCm and Intel OpenAPI, does not fall under this ban. This addition to the EULA was made in October 2021, marking a significant shift in policy.
Interestingly, around the same time in 2021, Intel withdrew its support for the Zluda project, signaling a lack of interest in enabling the launch of CUDA applications on Intel GPUs. Following this, the developer of Zluda, under the directive of AMD, began developing a similar layer for launching CUDA on AMD GPUs. However, AMD later withdrew its support for the project citing the same rationale as Intel. It is speculated that these changes in policies could be linked to the confrontation with Chinese GPU manufacturers who are working on a translating layer to run CUDA code on their chips.