Run CUDA on AMD hardware in Windows without virtualization

The battle for high-performance computing often feels like an exclusive club, neatly divided by proprietary standards. For years, the NVIDIA CUDA ecosystem has reigned supreme, locking developers into a single path for AI and deep learning workloads. But the landscape is shifting, particularly with AMD’s advancements and the latest ROCm updates providing robust official support for consumer GPUs like the Radeon RX 7000 and the upcoming RX 9000 series on Windows.

While AMD’s native ecosystem is catching up, a major hurdle remains: what happens when you need to run proprietary applications or specialized AI tools that stubbornly refuse to work without NVIDIA’s CUDA? This is where ingenuity steps in. A new open-source project is proving that running rigidly CUDA-exclusive workloads on AMD hardware in Windows is not just possible, but achievable without resorting to virtualization or dual-booting.

Enter the CUDA-for-AMD-Windows project, a clever toolkit designed to bridge this divide. It isn’t a simple runtime; it’s a highly automated and reproducible PowerShell setup that acts as a sophisticated translator between the worlds of ZLUDA—AMD’s translation layer—and AMD’s native ROCm SDK for Windows.

How does it work? Through a series of clever scripts, the toolkit automatically detects the user’s GPU architecture. It then seamlessly maps the necessary CUDA driver API, along with key libraries like cuBLAS, cuSPARSE, and cuFFT, directly over to their AMD equivalents. In essence, it intercepts the demands of CUDA and translates them into functional AMD calls, making it possible for developers to run CUDA-dependent software natively on AMD machines.

The proof-of-concept is compelling. The project successfully demonstrated end-to-end training of a 2.2-million-parameter PPO reinforcement-learning network using unmodified CUDA libraries on an AMD Radeon RX 9060 XT. This achievement shows that the translation pipeline functions effectively, proving that the barrier to entry for running CUDA-exclusive software on AMD hardware is primarily a translation tooling problem, not an insurmountable hardware flaw.

In controlled testing, the results highlight the efficiency of the official path. The “public upstream path,” relying purely on official ZLUDA releases and AMD’s HIP SDK, achieved a median throughput of 13,278 steps per second (SPS). Even when comparing this to a custom overlay built from salvaged legacy binaries, the official setup proved slightly faster, suggesting that while the translation process introduces minor overhead, the outcome remains highly competitive.

However, it is important to temper excitement with realistic expectations. This tool is a sophisticated hack, not an enterprise-grade deployment strategy. The author is transparent that it functions as a specialized adapter, meaning compatibility is strictly workload-dependent. Crucial components like cuDNN, TensorRT, and NCCL—which are vital for many advanced AI tools—do not yet resolve within this framework. Furthermore, the foundation upon which this project rests, ZLUDA, is currently being maintained as a “weekend hobby project”, introducing a level of risk for production work.

Despite these caveats, the work remains tremendously exciting. It proves that the path to open, flexible high-performance computing on AMD hardware is not blocked by corporate guardrails. With ongoing community contributions, the potential for expanding hardware detection and fixing stubborn library issues means this project could unlock a powerful new avenue for developers to harness the full potential of AMD’s growing GPU ecosystem.

#AMD#CUDA#NVIDIA#Radeon RX 7000#RX 9000 series

By Ethan Wright