CUDA-for-AMD-Windows Runs CUDA on Radeon RX 9060 XT

Developers who rely on NVIDIA's CUDA ecosystem can now run those workloads on AMD graphics cards in Windows without needing virtual machines or dual-boot setups. This shift matters because it removes the traditional hardware lock-in that forces AI researchers and hobbyists to buy NVIDIA GPUs for CUDA compatibility. A solo developer has created a toolchain that bridges the gap between AMD's native software stack and NVIDIA's dominant programming interface.

Solo developer Speedstu maps NVIDIA libraries to AMD hardware

The project, named CUDA-for-AMD-Windows, is the work of independent developer Speedstu. It functions as a translation layer that sits between CUDA applications and AMD's hardware. The tool combines the ZLUDA compatibility layer, specifically version v6-preview.69, with AMD's official HIP and ROCm software development kits. This combination allows Windows users to execute CUDA code directly on AMD silicon.

Tests on an AMD Radeon RX 9060 XT GPU show that the tool successfully maps several core NVIDIA libraries to their AMD equivalents. The system currently supports cuBLAS, cuSPARSE, and cuFFT, which are essential for linear algebra and signal processing tasks. However, critical machine learning libraries such as cuDNN, TensorRT, and NCCL do not resolve yet. This limitation means users cannot run complex deep learning inference or distributed training pipelines that depend on those specific packages.

Performance benchmarks on a 2.2 million-parameter reinforcement learning workload provide a baseline for the translation overhead. The public upstream path achieved a median throughput of 13,278 steps per second. A custom overlay version ran slightly slower at 12,876 steps per second, representing a roughly 3% performance drop. Speedstu noted that the repository is a tinkerer's tool rather than a corporate deployment strategy, highlighting its experimental nature.

ZLUDA itself is maintained as a weekend hobby project after losing commercial backing a second time. This status implies that the software may lack the rigorous stability testing found in enterprise-grade drivers. AMD recently updated its ROCm software to bring robust PyTorch and HIP SDK support to Windows for consumer GPUs, including the RX 7000 and RX 9000 series. This native support provides a stable foundation for the CUDA translation layer to operate upon.