AMD is set to release a Unified AI Software Stack by the end of the year, aiming to optimize AI workloads across CPUs, GPUs, and NPUs. This stack will utilize MLIR for intermediate representation and is expected to enhance AI workload management and developer experience. While details are still emerging, the stack promises to integrate with existing AMD software like ROCm and could potentially support non-AMD products.
AMD's ROCm now indirectly supports NVIDIA CUDA libraries through the open-source ZLUDA project, allowing AMD GPU owners to run CUDA apps within ROCm without code adaptation. ZLUDA, originally designed for NVIDIA CUDA support on Intel GPUs, has been successfully tested with AMD's Radeon GPUs on ROCm, enabling performance on par with NVIDIA for rendering tasks. While native CUDA support on Radeon GPUs remains unavailable, the potential of ZLUDA in bridging ROCm and CUDA stacks for AI applications is promising, with the developer exploring NVIDIA's upscaling capabilities on RDNA GPUs.
AMD quietly funded a project to create a drop-in replacement for CUDA on AMD GPUs, allowing many CUDA applications to run on the ROCm stack without modification. The project, based on the ZLUDA implementation, was developed by Andrzej Janik and is now open-source, dual-licensed under Apache 2.0 or MIT. While it's a promising solution, it's not without limitations, such as incomplete NVIDIA OptiX support. The implementation, leveraging the Rust programming language, has been tested successfully, and the exposed device name for Radeon GPUs via CUDA will be updated as part of the open-sourcing.
AMD is teasing open standards and open-source ahead of its "Advancing AI" event scheduled for December 6th. While details about the event are scarce, speculation suggests that it could involve the rollout of ROCm for more GPUs or the introduction of open-source/Linux support for Ryzen AI. AMD's promotion of open elements has sparked interest among enthusiasts, who are eager to see what announcements the event will bring.
AMD has announced that PyTorch machine learning development is now supported with the ROCm 5.7 series, specifically for the Radeon RX 7900 XTX and Radeon PRO W7900 GPUs. While it is unclear if other models are supported yet, AMD is actively working to bring ROCm to more RDNA 3 GPUs.
