ZLUDA Secures New Sponsor, Shifts Focus to AI/ML Workloads on Non-Nvidia GPUs
ZLUDA, an open-source CUDA translation layer, has undergone significant changes throughout its existence. Initially designed to run CUDA-based professional creative software on Intel and AMD GPUs, the project faced a major setback in August when AMD requested the takedown of its code. However, thanks to a new, anonymous sponsor, ZLUDA has been revived, this time with a focus on AI and machine learning (ML) workloads.
The new version of ZLUDA is designed to enable AI/ML software, such as PyTorch, TensorFlow, and Llama.cpp, to run on GPUs from multiple vendors, broadening its application beyond Nvidia’s CUDA ecosystem. Though the project will eventually support multiple GPU architectures, current efforts are concentrated on AMD’s RDNA1 and newer architectures. ZLUDA’s development is being built around AMD’s ROCm 6.1+ compute stack, paving the way for broader compatibility in the future.
According to Andrzej Janik, the project’s lead developer, the new ZLUDA code will require about a year to reach a stable state, but contributions from the open-source community are welcomed to accelerate progress. While ZLUDA will remain open-source for now, the identity of the mysterious sponsor remains undisclosed, sparking speculation that they are a major player in the AI/ML space. The sponsor’s decision to remain anonymous suggests they are not concerned about potential conflicts with Nvidia, which does not officially support using CUDA on non-Nvidia hardware.
The sponsor is expected to reveal their identity in the near future, which could provide more clarity about ZLUDA’s direction and the level of backing the project will receive going forward. For now, the community eagerly awaits the next chapter of ZLUDA, as it seeks to bridge the gap between CUDA and non-Nvidia GPUs in AI/ML workloads.
