The translation comes with a performance cost
In the quiet arithmetic of progress, AMD's newest upscaling technology, FSR 4.1, arrives bearing a paradox familiar to anyone who has watched innovation outpace its own installed base: the promise of better images delivered at the cost of slower performance for those running last year's flagship hardware. Testing by ComputerBase revealed that RDNA 3 GPUs — cards that were cutting-edge just twelve months ago — run FSR 4.1 measurably slower than its predecessor, a consequence of architectural choices made for a future those chips were not built to inhabit. The gap between generations is not merely marketing; it is written into the silicon itself, in the difference between how old and new hardware speak the language of artificial intelligence.
- Owners of AMD's RX 7900 XTX and other RDNA 3 cards are losing up to 14.5% of their frame rates simply by upgrading to the newest upscaling software — a penalty that arrives uninvited alongside a promised improvement.
- The culprit is architectural: RDNA 4 GPUs carry native FP8 AI accelerators, while RDNA 3 must translate that work through older INT8 integer pathways, creating friction that costs real frames in real games.
- AMD chose to prioritize visual parity across generations — ensuring the upscaled image looks the same on old and new hardware alike — but that design decision quietly transferred the cost onto the performance side of the ledger for older GPU owners.
- The situation is set to deepen: RDNA 2 cards, planned for FSR 4.1 support in early 2027, lack dedicated AI accelerators entirely and will rely on general-purpose stream processors, a slower workaround AMD has already acknowledged as more complex.
- The unresolved question hanging over all of it is whether the image quality improvement actually justifies the frame rate trade-off — a comparison that testing has not yet fully answered, leaving users to weigh a cost without a clear accounting of the benefit.
AMD's FSR 4.1 arrived promising better upscaling across its GPU lineup, but testing by German outlet ComputerBase uncovered an uncomfortable reality for owners of RDNA 3 hardware: the new software runs slower on last year's flagship cards than the version it replaced.
The numbers are hard to dismiss. On the Radeon RX 7900 XTX, FSR 4.1 in Quality mode dropped to 52.6 frames per second — an 11% regression versus FSR 3.1. Performance mode widened the gap to 14.5%. The RX 7800 XT and RX 7600 fared similarly, losing 7 to 9 percent depending on the mode. For gamers already pushing their hardware to its limits, these are meaningful losses.
The explanation lives in the architecture. AMD's newest RDNA 4 GPUs carry dedicated AI accelerators with native FP8 support — a modern shorthand for 8-bit floating-point math optimized for machine learning workloads. RDNA 3 cards have no such hardware and must approximate the same work through older INT8 integer calculations. AMD built FSR 4.1 to deliver visually identical results across both generations, but achieving that parity on older silicon required a different implementation — one that trades speed for consistency.
The trajectory points toward further strain. AMD plans to extend FSR 4.1 to RDNA 2 GPUs — the RX 6000 series — in early 2027. Those cards have no dedicated AI hardware at all, forcing the upscaling math onto general-purpose stream processors. AMD has acknowledged the added complexity, and while official performance figures remain unpublished, the pattern suggests RDNA 2 users could face steeper penalties still.
What remains unanswered is whether the visual improvement justifies the cost in frames. AMD's choice to prioritize image quality parity is coherent as a design philosophy, but it places a real burden on owners of hardware that was considered high-end just three years ago. The value of that trade-off depends entirely on what the upscaled image actually looks like — a question the benchmarks alone cannot settle.
AMD's newest upscaling technology, FSR 4.1, arrived with a promise: better image quality and performance across the company's GPU lineup. But when German tech outlet ComputerBase ran the numbers on RDNA 3 hardware—the generation that powered AMD's flagship cards just a year ago—they found something unexpected. The older chips were running the new software noticeably slower than the previous version, FSR 3.1.
The gap showed up clearly in the testing. On the Radeon RX 7900 XTX, FSR 4.1 in Quality mode delivered 52.6 frames per second, an 11% drop compared to FSR 3.1 running the same settings. Switch to Performance mode, and the penalty grew to 14.5%. The RX 7800 XT and RX 7600 told similar stories—roughly 7 to 9 percent slower depending on the mode. These aren't trivial numbers for gamers trying to squeeze every frame out of their hardware.
The reason traces back to how AMD built FSR 4.1 for different generations of chips. The company's newest RDNA 4 GPUs have dedicated AI accelerators with native support for 8-bit floating-point math, or FP8. RDNA 3 cards, by contrast, have to fall back on older 8-bit integer calculations, or INT8. That architectural difference forced AMD to implement FSR 4.1 differently on the older hardware, and the translation comes with a performance cost. AMD has said it aimed for visual parity across generations—meaning the upscaled image should look roughly the same whether you're running it on a 7900 XTX or a newer chip. That goal, it seems, came at the expense of speed on older silicon.
The situation will likely get worse before it gets better. AMD plans to bring FSR 4.1 to RDNA 2 GPUs—the RX 6000 series—sometime in early 2027. Those cards are even older and lack dedicated AI accelerators altogether. Instead, they'll have to use general-purpose stream processors to handle the upscaling math, a workaround that AMD acknowledged will be more complex and almost certainly slower. The company hasn't published performance numbers yet, but the trajectory suggests RDNA 2 users could face even steeper penalties.
For now, the gap between FSR 4.1 and FSR 3.1 on RDNA 3 raises a familiar tension in graphics technology: how much visual improvement justifies a performance hit, and whether that trade-off should be left to the user or baked into the software. AMD's decision to prioritize image quality parity across generations is defensible, but it also means owners of three-year-old GPUs are paying a real cost in frames per second to run the newest upscaling tech. Whether that cost is worth it depends on what the upscaled image actually looks like—a detail ComputerBase's testing didn't fully explore. That comparison may matter more than the raw frame rate numbers.
Citas Notables
AMD previously explained that there were notable differences between the implementation of FSR 4.1 on RDNA 4 and RDNA 3, specifically that the older hardware uses 8-bit integer data, while the RDNA 4 implementation uses built-in FP8 support— AMD, in interview with TechPowerUp
La Conversación del Hearth Otra perspectiva de la historia
Why would AMD release an upscaling tool that's slower on older hardware? Doesn't that defeat the purpose?
They're trying to solve two problems at once. FSR 4.1 is supposed to look better than FSR 3.1, but making it look the same across three generations of chips while also making it faster everywhere is nearly impossible. They chose image quality over speed on older cards.
So RDNA 3 owners are subsidizing the visual quality for everyone else?
In a way, yes. The newer RDNA 4 chips have hardware built specifically for this kind of math. RDNA 3 has to fake it with older instruction sets. AMD could have optimized FSR 4.1 separately for each generation, but then you'd have different-looking images depending on your GPU.
And RDNA 2 is going to be even worse?
Much worse, probably. Those cards don't have any AI accelerators at all. They'll be using general compute cores that weren't designed for this work. We won't know the numbers until early 2027, but the architecture gap is so large that double-digit performance hits wouldn't surprise anyone.
Does the image quality actually justify a 14% frame rate drop?
That's the question nobody's answered yet. ComputerBase measured the performance but didn't do a detailed visual comparison. You might not even notice the difference in a game, or it might be significant. Without that information, RDNA 3 owners are flying blind.