Andrzej Janik presented an update on the ZLUDA project, which develops an open implementation of CUDA technology. The aim of the project is to enable the execution of unmodified CUDA applications on non-NVIDIA GPU systems with performance close to that of applications running without layers. The project code is distributed under the MIT and Apache 2.0 licenses.
ZLUDA 4 is introduced as the first release formed after cleaning the codebase from code developed during Andrzej's time at AMD. From 2022, Andrzej worked at AMD on creating a compatibility layer for AMD GPUs with CUDA, but the project was discontinued in 2024. Following the contract's terms and after receiving publication permission from an AMD representative, Andrzej opened the code of the work done during his tenure at AMD, which allows CUDA applications to run on top of the ROCm stack and HIP runtime (Heterogeneous-computing Interface for Portability). In October, Andrzej was forced to remove the code from public access after a letter from lawyers indicated that the permission granted via email correspondence was not legally binding.
The new version of ZLUDA is based on the codebase that existed before Andrzej's work at AMD, is not tied to specific GPUs, and focuses on running applications that utilize CUDA to accelerate tasks related to machine learning and artificial intelligence. The project currently operates only on AMD GPUs, but will later be adapted for Intel GPUs. The release of ZLUDA 4 is written in Rust and is aimed at executing the Geekbench 5 test suite.
The new version includes a new parser for PTX, an intermediate language used in NVIDIA GPUs. This release is also notable for moving away from emulating NVIDIA GPU-specific increment and decrement instructions, instead calling similar AMD GPU instructions. These changes resulted in approximately a 10% performance increase during the Geekbench test compared to OpenCL and the ZLUDA 3 branch.
Source: opennet.ru
