With the release of ROCm 6.4.4, Radeon RX 9000 (RDNA 4), RX 7000 (RDNA 3), and the new Ryzen AI 300 “Strix” and Ryzen AI Max “Strix Halo” chips finally get the same treatment as their Linux cousins.

Until now, anyone wanting to run AI workloads on AMD kit under Windows had been left out in the cold while Nvidia’s CUDA monopoly thrived. ROCm 6.4.4 is only a preview release, but it at least means developers can start running models on Radeon hardware without installing Linux or resorting to hacks.

AMD is pitching this as part of its “cross-platform, developer-first” spiel. “Today we take an important step on that journey by enabling native PyTorch support on Windows and Linux for users with AMD Radeon 7000 & 9000 GPUs and Ryzen AI 300 & Ryzen AI Max APUs.” says Andrej Zdravkovic, Senior Vice President and Chief Software Officer

It took long enough, but now the kit you bought might actually do the job it was advertised for.

Developers can expect performance gaps and missing features for now, with AMD openly admitting this is more of a foundation to build on than a polished product. It’s an invitation for early adopters to bash the thing into shape while AMD promises improvements “as performance and feature coverage continue to improve and evolve.

On the enterprise side, AMD is already touting ROCm 7 on Instinct and EPYC hardware as the serious AI play, but consumer GPUs have always been left behind. This release is an attempt to make Radeons at least vaguely competitive in the AI developer ecosystem rather than just gaming cards with wasted compute potential.

Nvidia still has a stranglehold thanks to CUDA and developer inertia, but AMD now has a foot in the door into the Windows AI world. Whether developers actually walk through it or just shrug and stick with Nvidia remains the big question.