In the middle of August We shared here on the blog la news about AMD's request (demand) to Andrzej Janik, withdraw from the public domain a portion of the code for the ZLUDA project, a project that seeks to create an open implementation of CUDA technology.
Regarding the notification that AMD made at the time, this It came after six months after the code was made public, AMD lawyers contacted Andrzej, informing him that the permission granted during the correspondence had no legal validity.
You have to remember that The main goal of ZLUDA is to enable systems with non-NVIDIA GPUs run CUDA applications without modifications and with performance comparable to native applications, that is, without emulation layers that reduce their efficiency.
In 2022, Janik joined AMD to work on a compatibility layer CUDA for enterprise GPUs. However, After two years of development, AMD lost interest in this initiative. After initially receiving permission from an AMD representative to publish code developed during his work with the company, Andrzej released it to the public. However, six months after the publication, AMD lawyers contacted him, claiming that the permission granted was not legally valid. As a result, Janik was forced to remove ZLUDA code related to his work at AMD.
About this case, Andrzej Janik mentioned that he would not abandon the project and that he would look for an alternative route to continue development, but without having to resort to the advances he had made while working for AMD.
The code has been reverted to the pre-AMD state and I've been working furiously to improve the codebase. I've been writing the improved PTX parser I always wanted and have laid the groundwork for the rebuild.
And now It seems that the new starting point for the ZLUDA project has been generated, as Andrzej Janik recently presented the new plan for the future development of ZLUDA.
ZLUDA is back. For the past few months, I have been trying to find a commercial organization that will ensure the continued development of the project. I am happy to announce that I have found one that is not only willing to fund further development, but also has an excellent vision for the future of ZLUDA. I share their long-term vision and can't wait to talk more about it. We don't want to reveal everything yet, but for now, we know that we want to make ZLUDA better.
Andrzej Janik mentions that the new version of ZLUDA will be based on the original code developed before Andrzej Janik started working at AMD. This new implementation will not be tied to any specific GPU and will focus on running applications that use CUDA for machine learning and artificial intelligence tasks. Previously, ZLUDA focused primarily on content creation applications such as Arnold Render, Blender, and 3DF Zephyr.
In this new stage, ZLUDA will offer support for running frameworks such as Llama.cpp, PyTorch and TensorFlow with CUDA optimizations used by NVIDIA GPUs. Initially, the project will focus on support for AMD GPUs, and will later be adapted for Intel GPUs.
In addition to this, it is mentioned that the new ZLUDA will be designed to support multiple GPU architectures, with an initial focus on AMD GPUs. AMD's implementation It will be built from scratch, and is designed to be compatible with GPUs based on the RDNA1 architecture onwards.
Before the reversion of the above code, ZLUDA had been kept in ROCm 5, mainly to avoid retesting all the fixes specific to that version. Now that development is starting from scratch, the new ZLUDA implementation for AMD will use ROCm 6.1+, allowing it to take advantage of the most advanced features and improve compatibility on future architectures.
By Q2025 XNUMX, ZLUDA is expected to reach feature parity with the previously discontinued version, enabling AMD GPUs to run CUDA applications with competitive performance.
Finally If you are interested in knowing more about it, you can check the details in the following link