RealityScan 2.2 finally brings AMD GPU support (to Windows)

Epic have just released RealityScan 2.2, and the headline feature is one that’s been on people’s wishlists for a long time: full AMD GPU support on Windows.

For years, photogrammetry has been effectively CUDA-only. Most serious reconstruction software — Metashape, Colmap with GPU acceleration, the lot — all lean heavily on NVIDIA’s CUDA ecosystem. That means if you want to do photogrammetry at any scale, you’re locked into buying NVIDIA hardware. It’s been a genuine pain point for anyone running AMD GPUs, and it’s been frustrating because AMD has had solid OpenCL support for ages.

RealityScan has always been my favourite photogrammetry software (I wrote about 2.0 last year). It’s free (for hobbyists and companies under $1M revenue), it produces excellent results, and the integration with Unreal Engine is seamless. But until now, it was NVIDIA-only on Windows — which meant AMD users had to either buy an NVIDIA card or not use it at all.

What 2.2 brings

The new release brings hardware-accelerated reconstruction across every stage that previously required CUDA. The supported GPU list covers a reasonable spread of modern silicon:

– RDNA 3 (gfx1100): RX 7900 XTX/XT, PRO W7900/W7800

– RDNA 3 (gfx1101): RX 7800 XT, RX 7700 XT, PRO W7700

– RDNA 3 (gfx1102): RX 7600 XT / 7600 / 7650 GRE

– RDNA 3.5 (gfx1151): Ryzen AI Max “Strix Halo” APUs

– RDNA 4 (gfx1200): RX 9060 XT / 9060

– RDNA 4 (gfx1201): RX 9070 XT / 9070 / 9070 GRE, AI PRO R9700

That’s a solid range from the mid-range 7600 up to the workstation-grade W7900. The Strix Halo APUs are interesting too — means laptop users with those chips can finally get GPU-accelerated reconstruction without needing a discrete card.

There’s also a new feature worth mentioning: mixed AMD and NVIDIA setups will use both GPUs in parallel. If you’ve got a GeForce and a Radeon in the same machine, RealityScan splits the work across them. Excited to try that on my Asus Proart P16.

The Linux situation

RealityScan does have Linux support, but it’s worth being honest about what that actually means. The CLI works fine — you can run reconstruction pipelines headless without issues. But the GUI is essentially running inside a Wine container and has plenty of problems. I’ve tried using it myself and ran into issues with JXL image format support, which is annoying because JXL is one of the formats RealityScan handles well.

And this AMD GPU support? It’s Windows only. Linux users are on their own for now — Epic say Linux support is coming later but haven’t given a timeline. So if you’re running Linux and hoping to use an AMD GPU with RealityScan, you’ll need to wait or dual-boot.

At last

The broader point here is that photogrammetry has been stuck in CUDA land for so long it’s almost become accepted as normal. But the industry is shifting — AMD’s ROCm stack has matured significantly, and more software is recognising that not everyone wants (or can afford) an NVIDIA card. RealityScan being free already lowered the barrier to entry; now it doesn’t even require a specific GPU vendor.

Source: RealityScan 2.2 announcement