Andrzej Janik presented an update to the ZLUDA project, which develops an open implementation of CUDA technology. The goal of the project is to provide the ability to run unmodified CUDA applications on systems with GPUs other than NVIDIA GPUs, with performance close to that of applications running without interlayers. The project code is distributed under the MIT and Apache 2.0 licenses.
ZLUDA 4 is presented as the first release formed after cleaning the codebase from the code developed during Andrzej's work at AMD. Since 2022, Andrzej worked at AMD on creating a layer for AMD GPU compatibility with CUDA, but in 2024 the project was closed. In accordance with the terms of the contract and after receiving permission to publish from an AMD representative, Andrzej open-sourced the developments created during his work at AMD and allowing CUDA applications to be executed on top of the ROCm stack and runtime HIP (Heterogeneous-computing Interface for Portability). In October, Andrzej was forced to remove the code from open access after a letter from lawyers made it clear that the permission in the email correspondence was not legally binding.
The new ZLUDA release is based on the codebase that existed before Andrzej joined AMD, is not tied to any specific GPU, and is focused on running applications that use CUDA to accelerate tasks related to machine learning and artificial intelligence. The project is currently limited to running on AMD GPUs, but will be adapted for Intel GPUs in the future. ZLUDA 4 is written in Rust and is aimed at running the Geekbench 5 test suite.
The new version includes a new parser for PTX, the intermediate language used in NVIDIA GPUs. The release is also notable for moving away from emulating NVIDIA-specific modulo increment and decrement instructions, instead calling AMD GPU equivalent instructions. These changes resulted in about a 10% performance improvement in the Geekbench benchmark, compared to OpenCL and the ZLUDA 3 branch.
Source: opennet.ru
