Andrzej Janik) presented a plan for further development of the project zluda , which aims to create an open implementation of the technology cuda . The initial versions of Zluda focused on developing CUDA implementations for GPUs from Intel and AMD. However, the latest version aims to create a universal implementation of CUDA that can be used on any GPU other than NVIDIA’s. The goal of the Zluda project is to allow GPUs from different manufacturers to run unmodified CUDA applications with performance comparable to applications run on NVIDIA GPUs. The project code is available under the licenses of MIT and Apache 2.0.
Initially developed for Intel GPUs, the Zluda project faced challenges when Intel showed no interest in supporting the initiative to run CUDA applications on their GPUs. Subsequently, Andrzej left Intel and collaborated with AMD to develop compatibility with CUDA for AMD GPUs. However, after some time, AMD lost interest in supporting CUDA applications on their GPUs.
Following a contract with AMD and subsequent permissions, Andrzej opened up the code developed during his time at AMD to allow CUDA applications to run on AMD’s developed interface for portability, Hip . However, legal issues arose when AMD’s lawyers revoked the permission, leading to the removal of the Zluda code from open access.
The new version of Zluda will not be tied to a specific GPU and will focus on accelerating machine learning and artificial intelligence tasks using CUDA. This version will support applications like Llama.cpp, Pytorch, and Tensorflow with CUDA optimizations similar to those used for NVIDIA GPUs. Initially targeting AMD GPUs, the implementation will later be adapted for Intel GPUs, with plans to achieve full functionality parity by the third quarter of 2025.