It seems that AMD’s legal team intends to gain full control over a significant part of the software base created as part of the open-source ZLUDA project. According to The Register, earlier this year the company stopped financially supporting an initiative that allows CUDA code to be used on third-party accelerators. Now it looks like AMD is tightening its policy.
Initially, the ZLUDA project was created to run CUDA applications without any modifications on Intel GPUs with support from Intel itself. Later, the author of the project, Andrzej Janik, signed a contract with AMD, which included the creation of a similar tool for AMD accelerators. At the beginning of 2022, the project became closed, but already at the beginning of 2024, Janik made the project open again by agreement of the parties, since AMD decided to stop funding and further development of ZLUDA.
Image source: Mapbox/unsplash.com
However, AMD later changed its decision. It was at her request that the corresponding software became unavailable. According to Yanik, AMD’s lawyers stated that the previous letter of permission to publish the code was not a legally significant document. Janick, after consulting with a lawyer, concluded that the legality of the letters did not matter because a potential legal battle with AMD would take up too many of his resources and its outcome would be difficult to predict. It’s easier and faster to rewrite the project based on old developments, although some of the functions probably won’t be able to be recreated.
Why AMD decided to try to “bury” ZLUDA is not known for certain. The first and most obvious reason may be AMD’s desire to distance itself from a project that may violate NVIDIA’s intellectual property rights. NVIDIA has already prohibited the use of CUDA code on other hardware platforms by creating “CUDA translation layers” and resorting to decompiling everything created using the CUDA SDK to adapt the software to run on other GPUs.
In addition, AMD might have considered that the very existence of ZLUDA could prevent the implementation of its own software. AMD’s own tools involve porting and recompiling CUDA source code instead of running ready-made programs. Additionally, there could be a conflict regarding what code generated under ZLUDA can be released and what cannot.