AMD GPU Runs DLSS 5 at 119 FPS Thanks to Async Trick

A developer got DLSS 5's neural network running on an AMD RDNA 3 card, boosting frame rates from a rough 20 FPS to a smooth 119 FPS using an async approach.

Bet you never thought you'd see DLSS running on a red team card. Well, one clever developer just made it happen — and the numbers are wild.

A tech tinkerer going by RedDukeDev has dropped an experimental open-source project that gets DLSS 5's neural network model running on AMD graphics cards. The kicker? It takes performance from a basically unplayable 20 FPS all the way up to 119 FPS.

Here's the problem the project had to solve. DLSS 5's reconstruction work is seriously heavy, and third-party GPUs like AMD's don't have dedicated Tensor Cores. Without that specialized hardware, the neural network module introduces some pretty painful processing latency.

If you go with the standard synchronous approach — where every single frame has to wait for the neural network to finish its calculations — performance tanks to a level most of us would call unplayable. To prove the point, the developer ran a benchmark on an RDNA 3 card at 1440p, with ray tracing off and FSR 4 upscaling plus frame generation turned on.

The baseline scene was pumping out around 139 FPS. Switch to synchronous neural network computation, though, and it crashes down to roughly 20 FPS. Yeah, that's rough.

The async method flips the script. Instead of making the display pipeline sit around waiting, the neural network's output gets processed the moment it's ready. The system calculates the residual between the network's input and output, then applies the result to the current frame using a displacement map. That way, the display pipeline never has to pause, and frame rates hold steady at 119 FPS — keeping most of that original smoothness intact.

It's not all sunshine, though. The approach still has some clear limitations right now. In dynamic scenes, you'll spot visual artifacts, and they get worse the bigger the time gap between rendering and neural network computation, especially with frame generation enabled. In static scenes, though, the distortion is basically nonexistent.

The developer is already working on adding various optimization algorithms to cut down on ghosting, and has confirmed the tool plays nicely with Linux and the Proton compatibility layer. Not bad for an experimental side project — and a pretty interesting peek at what's possible when you get creative with hardware that wasn't built for the job.

Would you risk a few artifacts to run DLSS on an AMD card? Let us know what you think.