I posted a small snippet of information on bluesky recently about reviewers comparing upscalers. I have seen outlets compare images rendered with GPUs in different market segments, such as AMD Radeon RX 9070 XT to Nvidia Geforce RTX 5090. I stated if this is done, FPS needs to be limited to be a fair comparison of upscale quality - this post explains the technical reason as to why.

It’s in the name

These modern upscalers all replace Temporal Anti Aliasing(TAA) in the render pipeline. TAA aims to anti-alias images by spreading the expensive operations of super sampling the input signal over multiple frames, rather than techniques such as MSAA which selectively super sample within the same frame spatially.

Just knowing this begins to expose why image quality can be tied to FPS - more frames directly equals more samples of the input signal to derive a high quality output image.

This can be further expanded with some knowledge of how TAA and upscalers such as FSR 3 work. The super-sampling of the input over multiple frames allows for anti-aliasing of edges and bringing out fine detail because every view is offset by a sub-pixel amount each frame. If TAA was not operating on the rendered output, your game would wobble about and be very unstable.

The amount of sub-pixel jitter applied is a well defined sequence pattern. There are many to choose from, but FSR and other upscalers recommend a Halton sequence. You can see in the FSR documentation that this sequence has a specific length, depending on the upscale ratio - the larger the upscale, the longer the sequence.

For a 2x upscale, the sequence length that is recommended for FSR and XeSS is 32. That means 32 frames are required to complete the sequence, and fully converge to a super-sampled output pixel. You may think this sounds very long - and it is. The algorithm utilizes motion vectors and other tricks to keep samples valid so previous data can be used throughout the convergence. You can read more about the inside workings of TAA via this article by Emilio López.

How it shows

In early temporal FSR, output pixels which had not seen a full convergence (can also be referred to as accumulation) through this sequence could appear “fizzy” - this is the jitter sequence resolving in front of your eyes. Disocclusions and thin features where convergence is “reset” are particularly sensitive to this, as was seen in many reviews. There are ways to reduce how this appears, such as de-jitter blur, but fundamentally if an output pixel does not have a full jitter sequence of samples “input” to it, then it is likely the output quality would be lower due to missing sample data.

So, back to the original statement: higher FPS allows for more fully accumulated and converged output pixels, directly affecting output quality. But, by how much?

Quantifying the problems

Finding benchmark data for FSR 4 vs DLSS 4 at the same modes had proven tricky. However, if we grab 1440p average game data for the GPUs in question from Hardware Unboxed as a ballpark figure, we arrive at Radeon RX 9070 XT at 119fps versus Geforce RTX 5090 at 192fps.

If our jitter sequence is 32 frames for a 2x upscale, that means there are 3.71 full accumulations per second capable on Radeon RX 9070 XT, or 6 on Geforce RTX 5090.

In other terms, at these frame rates it takes 269ms to fully accumulate on Radeon RX 9070 XT, or 166ms on Geforce RTX 5090. A huge 100+ms difference in convergence rate. With animation, particles and then camera motion in play during these quality comparisons, that is a lot of time to display lower quality pixels.

These days we are all getting told anti lag improvements of 15-20ms are groundbreaking. That frame generation latency hits of 30ms are unacceptable. So should a graphical convergence difference 2-3 times that be acceptable during a quality review?

Conclusion

Different upscalers all have different considerations. TSR used in Unreal Engine specifies its convergence rate as a function of frame rate. The main takeaway is that higher FPS provides better upscale/TAA quality.

Graphical comparisons of these upscaling techniques should be done, where possible, at the same fps. If you still want to go for the market competition angle, at least compare using the FPS capable on the competing card matchups. Don’t run a £620 card against a £2000 card.

This becomes more of an issue now with FSR not running cross-platform, which complicates things. Previously folks could run FSR 1-3 on the same RTX card that was used to capture DLSS - all at around the same level of base FPS.

That’s it for this one. Any comments or questions can be directed to me on Bluesky @domipheus.

---

Many thanks to these lovely folks who took time to review my post:

• Rys Sommefeldt
• Neil Henning
• Luke Iwanski