Project zluda has developed an open implementation of CUDA technology for AMD GPUs, enabling the execution of unmodified CUDA applications with performance comparable to applications executed without layers. The project provides tools that ensure binary compatibility with existing CUDA applications compiled using the CUDA compiler for NVIDIA GPUs. This implementation is built on top of AMD’s Steki rocm and the HIP runtime (Heterogeneous-computing interface for portability). The project’s code is written in Rust and is licensed under MIT and Apache 2.0. The implementation is supported on Linux and Windows.
The development of a CUDA compatibility layer for AMD GPU systems has been ongoing for the past two years. However, the project has a longer history and was initially created to enable CUDA functionality on Intel GPUs. The change in GPU support is due to the fact that the original developer of Zluda was an Intel employee. However, in 2021, Intel deemed the launch of CUDA applications on Intel GPUs as not economically viable, which led to the discontinuation of support for Zluda on Intel GPUs.
In early 2022, the developer left Intel and entered into a contract with AMD to develop a compatibility layer for CUDA. As part of the contract, AMD requested that the developer refrain from advertising AMD’s involvement in the Zluda project and from making commits to the public Zluda repository. Two years later, AMD concluded that launching CUDA applications on AMD GPUs is not aligned with their business objectives. As a result, the developer