Andrzej Janik presented the updated project zluda, openly developing an implementation of the cuda technology. The main goal of the project is to enable unmodified CUDA applications to run on GPU systems other than NVIDIA with performance close to native applications. The project code is distributed under the MIT and Apache 2.0 licenses.
zluda 4 is the first release after cleaning the code base of code developed during Andrzej’s time at AMD. He worked at AMD from 2022 to create a compatibility layer for AMD GPUs with CUDA, but the project was discontinued in 2024. After obtaining permission from an AMD representative, Andrzej published the code developed during his time at AMD, enabling CUDA applications on the rocm stack and HIP runtime. However, the code had to be removed from open access in October due to legal reasons.
The new version of Zluda is based on the code developed before Andrzej joined AMD, is not tied to specific GPUs, and focuses on accelerating machine learning and artificial intelligence tasks using CUDA. Although currently limited to AMD GPUs, it will be adapted for Intel GPUs in the future. Zluda 4, written in Rust, is specifically designed for running tests with geekbench 5.
The new release includes a Parser for ptx, the intermediate language used in NVIDIA GPUs. It also emulates AMD GPU instructions, resulting in a 10% performance increase in geekbench testing compared to OpenCl and Zluda 3.
