Input latency is the all-too-frequently missing piece of framegen-enhanced gaming performance analysis

We live in rapidly changing times for PC graphics. More and more pixels are being generated by AI or “neural” rendering techniques. The DLSS, FSR, and XeSS upscalers let us boost frame rates at will in exchange for small dings to image quality. Frame generation techniques, especially Nvidia’s multi-frame generation tech on its Blackwell cards, let us boost them even higher still with the click of an in-game toggle.

When implemented well, we think MFG can be an improvement to the gameplay experience. Take crushing titles like Alan Wake II and Cyberpunk 2077. On a high-refresh-rate, variable-refresh rate monitor, both games look pretty choppy at 45 or 50 FPS if you’re running them at demanding settings, even with an assist from a monitor's low framerate compensation.

Blackwell’s multi-frame generation can turn that choppy ride into a smooth experience by pushing output frame rates far higher up in a display’s refresh rate range. There are some cases where MFG’s seams are still visible, but we think they’re an acceptable tradeoff for the much smoother output.

There’s just one catch: a game with MFG can look smoother, but it might not feel better to interact with. That’s because its input lag remains tied to the speed at which the base game loop runs. DLSS and other upscalers can improve this latency by boosting baseline performance, but framegen and MFG don’t.

We’ve frequently, albeit informally, described this need for a good input lag baseline in past coverage. (See also our ruminations on framegen performance in Black Myth Wukong.) As framegen and MFG both seem poised to become a bigger and bigger part of PC gaming going forward, we think that there needs to be a more formal, empirical way of communicating this latency baseline.

Having an easy way to quantify both input lag and output frame rate lets us communicate better with everybody about what we're seeing, and exploring the relationship between baseline frame rate, input lag, and output frame rates with framegen can also help us work back from manufacturer info to determine whether claims were made with an acceptable input latency as a baseline.

We already have the tools to at least approximately capture this latency. Nvidia’s own FrameView application provides a “PC Latency” metric, while Intel’s PresentMon exposes a measurement called “All Input to Photon Latency,” which can be optionally captured as part of its overall performance logging.

We've found game and GPU vendor compatibility potholes with both of these applications, and we have some reservations about these entirely software-based metrics, but higher latency measurements from them correlate to a notably worse-feeling gameplay experience, and that’s good enough as a starting point.

Most importantly, communicating this measurement alongside framegen-enhanced output makes for more honest marketing. Nvidia apparently knows this, since it touted low PC latency as an advantage in Cyberpunk 2077 when it introduced MFG alongside the RTX 5090. In contrast, it’s conspicuously absent from the recently-launched RTX 5050’s MFG-heavy official marketing materials.

In any case, if a graphics card isn't contributing to a good input-to-photon latency to begin with, pumping its output frame rate with framegen or MFG—and only communicating that figure—doesn’t show the complete picture.

We wanted to see if quantifying this latency provides a more grounded description of what it’s like to game in a future where output frame rates might be less about raw power and more about toggling in-game settings to taste.

Put your cards on the table

Our not-particularly-original proposal here, but one that we need to formalize regardless, is that there is a maximum acceptable latency threshold with framgen enabled (call it the MALT) beyond which a given game starts to feel more and more uncoupled from user input. This unpleasant behavior presents itself in various ways, like "rubber banding" or "swimminess."

When talking about framegen performance, we believe companies and testers should include the MALT they considered “acceptable” when using those techniques to boost the output frame rate.

Lower MALTs will require higher baseline performance, whether through beefier hardware or lighter in-game settings. And there is a ceiling past which a MALT will be too high to ensure a good interactive experience.

What makes for a good MALT? It’s likely to vary by game and by game type. Our experience with Alan Wake II and MFG so far suggests that any input-to-photon latency below 30 ms is excellent, 30-45 ms feels great, 45ms - 60 ms is usable, and anything too far above that will start to cause perceptible rubber banding or discontinuity between input and output.

Why measure this latency directly instead of just using frame rates as a proxy (i.e., 60 FPS as a “good enough” foundation on which to enable MFG?) It’s probably an edge case, but Nvidia has highlighted at least one experience where a frame rate that intuitively seems “good enough” does not match up with an acceptable input latency. The company observed Microsoft Flight Simulator running at 50 FPS natively, but with an input latency of 172 ms.

In short: if you can measure it, measure it.

Achievable latency thresholds are also likely to vary from system to system, due to the huge array of peripherals, components, and monitors out there. Your tolerance for input lag may also be tighter or looser than ours. PresentMon and FrameView are free, so the only cost involved in testing this out for yourself is time.

Our testbed for this article uses the following components:

Exploring MALT-normalized performance

In a world where even the $329-ish RTX 5060 Ti 16GB can push 200 FPS at 1440p with DLSS and MFG 4X in Alan Wake II, we wanted to see what performance analysis might look like if we followed our own advice and normalized to a MALT. We picked 60 ms as our threshold and got to work seeing how we could tune settings on a range of modern GPUs to stay within that range.

Our results are fairly intuitive. The first column shows the system latency you get simply from moving up to a more powerful graphics card with the same settings, and the second then shows the higher settings we’re able to select within that headroom without going over our 60 ms input latency threshold. Read the chart in a Z pattern to understand the enhancements that a more powerful card gives you for the same MALT.

Buying a more powerful GPU for Alan Wake II, at least, intuitively lets you enjoy lower input latency and higher output frame rates than those at our 60 ms MALT for a given group of settings, or higher resolution and image quality settings for the same 60 ms threshold.

The point here is that if you don’t need the lowest possible input latency, as you might not in a slower-paced experience like Alan Wake II, putting more of your GPU’s resources toward higher-resolution output and richer visuals when you have the latency headroom to do so with framegen is just smart. Monitoring this figure also lets you dial back settings if your given choices result in an unpleasant experience.

One challenge in fine-tuning performance with MFG is that DLSS quality settings only offer us five coarse levels of adjustment today without the use of third-party tools. If we had more intermediate DLSS settings, we could potentially achieve lower input latencies (imagine a value between Ultra Performance and Performance, for example) than our 60 ms threshold while still maximizing image quality within the rendering resources of a given graphics card. We hope that Nvidia explores ways to natively expose finer DLSS granularity in the future to make the most of MFG.

MALTed marketing

Considering latency and frame rates together also helps us analyze a bold question: does MFG let lower-end GeForce cards “beat” their higher-end predecessors? Nvidia infamously claimed the RTX 5070 could beat the RTX 4090 with the full suite of DLSS 4 features enabled. We didn’t test that exact claim, but we did pit the RTX 5080 versus the RTX 4090 to try and understand whether it had any merit.

If we normalize for a 60 ms MALT in Alan Wake II with high raster settings, ultra RT, and a 4K target resolution, the RTX 5080 has to use DLSS Ultra Performance to achieve this target input latency, whereas the Ada flagship can use DLSS Performance for a boost in image quality at the same latency.

The major difference, of course, is that MFG 4X lets the RTX 5080 output an enhanced 190 FPS with DLSS Ultra Performance, for a theoretically much smoother output than the RTX 4090’s 97 FPS with DLSS Performance. The difference in delivered smoothness isn’t as drastic to our eyes as the output frame rates might suggest, but our 160 Hz monitor might be concealing the full magnitude of this improvement.

Use DLSS Ultra Performance on both cards, though, and the RTX 4090’s PC latency drops 25%, to 46 ms on average, and the output frame rate rises to 133 FPS. Even though the RTX 4090’s 2X framegen output still can’t match the RTX 5080’s with MFG 4X at the same settings, we enjoy much lower overall input latency on the RTX 4090 than on the RTX 5080, and some might prefer that more responsive experience.

Looking at both input latency and output frame rates together, the RTX 5080 can deliver theoretically smoother output, but RTX 4090 owners probably shouldn’t go out and upgrade on this basis alone. And it certainly throws cold water on bold claims that even lower-end cards than the RTX 5080 can match or exceed the gaming experience of an RTX 4090 unless the output frame rate is all that you’re measuring.

This leads us to Nvidia’s marketing materials for the recently launched RTX 5050. Eyeballing its charts, the company appears to have used a low baseline frame rate for the non-esports titles in its selection. Cyberpunk 2077 and Avowed both appear to cap out around 160 FPS with MFG 4X, implying a baseline of around 40 FPS on average.

Again, we’ve already seen that intuition doesn’t necessarily serve us well when tying frame rates to a particular game’s input latency, and it’s entirely possible that Nvidia observed an acceptable latency threshold when it tuned settings to achieve that 40-ish FPS. But when you appear to be riding the line like this, it makes disclosing those assumptions all the more important.

Intel recommends a minimum of 40 FPS for use with its Xe Frame Generation tech, but that figure is for developers targeting a good experience on integrated graphics. For higher-end graphics cards, Intel suggests a 60 FPS baseline is ideal. AMD recommends 60 FPS as a performance target before using FSR 3 Frame Generation, period (and also cautions that “sub 30fps frame rate pre-interpolation should be absolutely avoided.”)

Nvidia doesn’t have a recommended frame rate for use with MFG that we could find, but it’s also got a trick up its sleeve that Intel and AMD’s methods lack. Nvidia Reflex in DLSS 4 includes Frame Warp, a reprojection technique that we already know from VR. Frame Warp uses a range of data sources to shift output frames to more closely match the user’s most recent inputs, reducing perceived latency.

It’s possible that with Frame Warp and the perceptual latency reduction it purports to offer, MFG can tolerate input frame rates lower than XeFG and FSR—but again, these assumptions can be measured, and they should be communicated. After all, if Nvidia’s framegen solution can robustly tolerate a higher input latency and do more with less on lower-end hardware like the RTX 5050, that’s a competitive advantage over other framegen approaches.

The bottom line - for now

With all this in mind, how should we reason about GPU performance in a world where framegen is more and more common? MFG does not suddenly mean that more powerful GPUs are unnecessary or that lower-end parts can magically eclipse higher-end ones. It’s more of a cherry on top of the power you’re already buying than a leg up on the next card up the stack.

Higher-end GPUs can still give you better image quality at the same input latency or a much lower input latency for the same settings, and you still need their muscle to push demanding games at higher output resolutions, period, even with upscaling in the picture.

What’s also important to remember is that not all framegen techniques are the same, even from the same vendor. Framegen on Nvidia Ada cards looks a bit worse to our eyes than on Blackwell with MFG, and MFG in DLSS 4 includes technologies like reprojection that other vendors’ solutions lack. We didn't test FSR or XeSS frame generation for this article, but they lack the flexibility that MFG's 3X and 4X multipliers give you for better matching output frame rate to your monitor's frame rate.

On that note, we think the best way to think of MFG is as a bridge to increasing gaming smoothness on today’s high-refresh-rate monitors, even with relatively attainable graphics cards. For just one example, 1440p 240Hz monitors used to be wildly expensive, but you can now grab one for less than $300. The cost of GPU power is obviously not going down at the same rate as those high-refresh-rate monitors.

Recall that the $329-and-up RTX 5060 Ti 16GB can push 200 FPS at 1440p with DLSS Ultra Performance and MFG 4X in Alan Wake II. To reach those same performance figures without MFG at the same settings, you need to use a $3000+ RTX 5090. We know, because we tried.

Unless you’re an esports pro who needs both the lowest possible input latency and high output frame rates, being able to use higher reaches of your monitor’s refresh rate range without spending $3000 on a commensurate GPU is pretty cool.

The worst way to think about MFG is as a route to achieving “playable” output frame rates. This is a false grail. A game running at an unplayable 20 FPS pre-framegen cannot magically be boosted to a “playable” 60 or 80 FPS using MFG, simply because the input latency at such a low native frame rate is likely to exceed any acceptable standard.

Overall, our brief investigation suggests that in a future where framegen is more and more common, both vendors and media alike need to be clearer about the baseline input latency they’re assuming when talking about framegen-enhanced performance. It’s entirely possible for all camps to end up creating marketing materials and performance or image quality analyses that rest on an input frame rate that’s too low and an input latency that’s too high for the end user to actually enjoy.