The user, going by the handle “Virtual-Cobbler-9930,” managed to get FSR 4 running unofficially, and shared benchmarks that provide a visual upgrade that comes with a hefty speed penalty, making it a rather mixed bag for RDNA 3 owners.
In Cyberpunk 2077 at 4K using the FSR Quality setting, FSR 4 managed 56.3 frames per second. FSR 3.1 clocked a much brisker 85.1. The visuals with FSR 4 were undeniably sharper, but it turns out prettier doesn’t mean faster.
In Oblivion Remastered, FSR 4 helped clean up image artifacts that plagued FSR 3.1, but the price was a 20 to 30 per cent hit to performance. You can claw back some frames by dialling things down to “Balanced” mode or using Frame Generation, but it’s still not ideal.
Esports title Marvel’s Rivals was even more brutal. FSR 4 dropped frame rates to a sluggish 41 FPS, compared to 74 FPS with FSR 3.1. For fast-twitch gamers, that’s a deal-breaker. One Redditor said: “That gain could be almost nullified at lower resolutions like 1080p. Simply because RDNA3 can’t push higher FPS with FSR4 due to architecture.”
The issue lies with how FSR 4 leans heavily on AI and machine learning. RDNA 4 has the hardware to handle that. RDNA 3 doesn’t. Without dedicated AI cores, FSR 4 runs like it's dragging a sack of bricks uphill. That forces it into a slower, clunkier path to upscale your visuals.
Strategically, AMD’s move is a bit cheeky. By letting RDNA 3 users fiddle with FSR 4 unofficially, it builds hype without officially endorsing it. It subtly tells the world that if you want the real deal, you’ll need RDNA 4. It’s a tech teaser masquerading as backward compatibility.
If you’re chasing maximum image quality at 4K and don’t mind tinkering in your spare time, FSR 4 on RDNA 3 might scratch the itch. But for anyone playing esports titles or anything twitchy, you’ll get better mileage from sticking with FSR 3.1 or just running native.
Right now, FSR 4 on RDNA 3 is a diplomatic gesture, not a viable solution. It shows what’s possible, but also where AMD drew the performance line.
