{"id":120066,"date":"2024-10-07T01:09:14","date_gmt":"2024-10-06T23:09:14","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/proekt-zluda-prodolzhit-razvitie-v-forme-universalnoj-realizaczii-cuda"},"modified":"2024-10-07T01:09:14","modified_gmt":"2024-10-06T23:09:14","slug":"proekt-zluda-prodolzhit-razvitie-v-forme-universalnoj-realizaczii-cuda","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/news\/proekt-zluda-prodolzhit-razvitie-v-forme-universalnoj-realizaczii-cuda","title":{"rendered":"The ZLUDA project will continue its development as a universal implementation of CUDA.","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Andrzej Janik presented the plan for the further development of the ZLUDA project, which is advancing an open implementation of CUDA technology. While the first version of ZLUDA aimed to create a CUDA implementation for Intel GPUs, and the second one for AMD GPUs, the currently presented third version will focus on developing a universal CUDA implementation for any GPUs other than NVIDIA GPUs. The goal of the ZLUDA project is to provide systems with non-NVIDIA GPUs the ability to run unmodified CUDA applications with performance close to that of applications running without layers. The project's code continues to be distributed under MIT and Apache 2.0 licenses. <\/p>\n
Initially, the ZLUDA project was developed for Intel GPUs, but in 2021, the company considered the ability to run CUDA applications on Intel GPUs to be of no business interest and ceased supporting the development of the initiative. In early 2022, the developer left Intel and entered into a contract with AMD to develop a compatibility layer for CUDA for the company\u2019s GPUs. After two years, AMD lost interest in the ability to run CUDA applications on AMD GPUs. <\/p>\n
In accordance with the terms of the contract and after obtaining permission to publish from an AMD representative during email correspondence, Andrzej opened the code developed during his time at AMD, which allows running CUDA applications on top of the AMD-developed ROCm stack and HIP (Heterogeneous-computing Interface for Portability) runtime. Six months after publication, AMD's lawyers reached out to Andrzej and indicated that the previously granted permission during correspondence was not legally binding. As a result, Andrzej was forced to remove the ZLUDA code created during his time at AMD from public access. <\/p>\n
The new version of ZLUDA will be based on the codebase that existed before Andrzej started working at AMD, will not be tied to any specific GPUs, and will focus on executing applications that use CUDA for accelerating machine learning and artificial intelligence tasks (formerly, ZLUDA was mainly geared towards content creation applications, such as Arnold Render, Blender, and 3DF Zephyr). Among other things, it plans to provide support for Llama.cpp, PyTorch, and TensorFlow frameworks on top of ZLUDA utilizing CUDA optimizations applicable for NVIDIA GPUs. <\/p>\n
Initially, the project will focus on running on AMD GPUs, but it will later be adapted for Intel GPUs. The implementation for AMD GPUs will be built from scratch, will support GPUs based at least on the RDNA1 architecture, and will use the ROCm 6.1+ stack instead of the ROCm 5 stack that was used in the earlier AMD implementation. Parity of functionality with the previous implementation, which Andrzej had to remove from public access, is planned to be achieved by the third quarter of 2025. \n Source: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"