奕然 发自 凹非寺 量子位 | 公众号 QbitAI 能让非NVIDIA芯片跑CUDA的开源项目ZLUDA，起死回生了。 最新版增加了对大模型工作负载的支持，一举登上GitHub热榜！ 开发者@vosen本名Andrzej Janik，曾经在Intel工作。 该项目一度因AMD停止资助濒临破产，最终被一家神秘机构出手相救。 现在，创始人vosen带来好消息，表示ZLUDA团队新添一员猛将并稳定进行项目恢复中。 在2020年，Andrzej Janik想要尝试一下技术突破，让CUDA程序在非NVIDIA平台运行，一尝试，便有了可行性。 之后，ZLUDA被Intel接手，作为一个内部试验项目发展。 Intel分配资源给了ZLUDA，目的很明显了，让其 在Intel GPU上跑CUDA程序，作为Intel oneAPI生态的一种补充方式 。 无疑，这触碰到了NVIDIA的商业生态链。 没过多久，这个项目就被终止了。 2022年，ZLUDA得到了AMD的支持而重启，并支持AMD硬件。 好景不长，这次也仅仅维持2年，2024年2月宣布终止。 过往发展：起起伏伏又起起 一个月后，英伟达就发布CUDA 11.6版本，并明确 ...

Core Viewpoint - The article discusses the advancements of the open-source project Zluda, which aims to enable CUDA applications to run on non-Nvidia GPUs, thereby expanding hardware options and reducing vendor lock-in [4][7]. Group 1: Zluda Project Updates - Zluda has made significant progress in achieving CUDA compatibility on AMD, Intel, and other third-party GPUs, allowing users to run CUDA-based applications with near-native performance [4][7]. - The team behind Zluda has doubled in size, now including two full-time developers, which is expected to accelerate the project's development [4]. - Recent updates include improvements to the ROCm/HIP GPU runtime, ensuring reliable operation on both Linux and Windows platforms [5]. Group 2: Performance Enhancements - The performance of executing unmodified CUDA binaries on non-Nvidia GPUs has significantly improved, with the tool now capable of handling complex instructions with full precision [7]. - Zluda has enhanced its logging capabilities to track interactions between code and APIs, capturing previously ignored interactions and intermediate API calls [7]. - The project has made notable progress in supporting llm.c, a pure CUDA test implementation for language models like GPT-2 and GPT-3, with 16 out of 44 functions implemented [7]. Group 3: 32-bit PhysX Support - Zluda has received minor updates related to 32-bit PhysX support, focusing on efficient CUDA log collection to identify potential errors that may also affect 64-bit PhysX code [8]. - Full support for 32-bit PhysX may require significant contributions from third-party developers, indicating a collaborative effort is needed for further advancements [8].