It’s hard to say if the RTX 5090 or DLSS 4 won Nvidia more limelight over the past week. The RTX 5090 is undoubtedly the best graphics card you can buy, but a big part of that conversation has involved DLSS 4 and its multi-frame generation capabilities. And for good reason. DLSS 4 is remarkable, enabling gaming experiences like full path tracing in Alan Wake 2 at triple-digit frame rates without breaking a sweat.
Still, I’m worried.
DLSS 4 represents a radical change for all RTX graphics cards, from multi-frame generation for the latest RTX 50-series GPUs to a new AI model that reaches down to cards like the RTX 2060 Super. For as impressive as DLSS 4 is, however, I’m concerned that the floor will fall out from under it eventually.
Get your weekly teardown of the tech behind PC gaming
Transforming the DLSS model
DLSS 4 is a much more significant update than you might believe. The headlining feature is DLSS Multi-Frame Generation or MFG, but this new version has tweaked every aspect of DLSS. That means that, yes, DLSS 4 will retroactively make DLSS better, even for older GPUs that don’t support MFG.
The big change is that Nvidia went from a convolutional neural network, or CNN, to a transformer model. Some games, such as Cyberpunk 2077 and Alan Wake 2, let you switch between the different models within the game. However, I suspect most games with DLSS will just stick with the transformer model moving forward. If, for some reason, you prefer the CNN, you can revert to it through the Nvidia app with its new DLSS override feature.
Nvidia says that the new transformer model has more parameters, but more importantly, transformer models are self-referential. Instead of tracking every pixel throughout the scene, the new DLSS model can track specific pixels to improve image quality in problematic areas. Nvidia has demonstrated this capability, which you can see in the video above. There are clear improvements, but they aren’t universal.
Above, you can see a comparison between the CNN and transformer model in Cyberpunk 2077. Without labels — or even with them, actually — it’s basically impossible to tell the difference between the two images. I searched and searched, peeping at every pixel, to find some sort of clear difference between the two, and it’s just not there. Both look great, especially considering I’m running in DLSS’ Performance mode, but I wouldn’t say the transformer model is clearly better.
The situation is identical in Alan Wake 2, where, once again, it’s basically impossible to tell the difference between the two images. There’s a slight difference in the detail on the bench, but I’m struggling to attribute that to DLSS when there’s some layer of water blurring the camera.
But here’s the thing — the new transformer model doesn’t need to be better, especially right now. What’s important is that it’s just as good as the CNN model, at minimum, and with no hit to performance. You’ll see the benefits in some games and situations, but the transformer model isn’t a glossy new coat of paint that covers all of DLSS. Treat it like one of Nvidia’s smaller DLSS updates, which have quietly made the feature better over the past several years.
There are some improvements here, too. Take a look at another scene from Cyberpunk 2077 above. The new transformer model not only reproduces the reflections more clearly, it also deals with a ton of ghosting in the CNN model. There are these long, obvious trails behind each tail light as the cars pass the camera with the CNN model, and they completely disappear with the transformer model.
This will likely improve over time. Nvidia has said that it has a supercomputer that’s been improving DLSS for the past six years, and that work isn’t stopping with the new transformer model. Although the benefits might not show up everywhere, they’re still a welcome addition. The transformer model is available to all RTX GPUs, and it covers not only DLSS Super Resolution but also Ray Reconstruction.
One frame, two frame
OK, but I know why you’re all here. Let’s talk about MFG. Exclusive to RTX 50-series GPUs like the RTX 5090, DLSS 4 unlocks up to 4X frame generation — one rendered frame alongside three generated ones. There are 75 confirmed games that’ll support DLSS 4, though not all natively. Some will have MFG options in the graphics menu, while the rest will need a DLSS Override from the Nvidia app.
And MFG is really good in DLSS 4. The chart above speaks for itself on that front. In Alan Wake 2, I went from less than 30 frames per second (fps) on the RTX 5090 to well over 200 fps. That’s an exponential increase in performance and with excellent image quality in tow. Latency isn’t an issue, either. Nvidia does a wonderful job these days of limiting the latency hit with frame generation, and adding a couple of extra frames into the mix doesn’t significantly increase the latency — it barely moves the needle.
The situation is the same in Cyberpunk 2077. There’s barely any additional latency with frame generation turned on, even up to 4X mode. Add DLSS Super Resolution into the mix to increase the base frame rate — and reduce the total latency — and you’re sitting near 240 fps at 4K. The superpower of MFG is that it’s multiplicative; if you feed it with a higher base frame rate, you’ll get a larger performance boost out of it.
You’ll get the best experience out of MFG if you feed it with at least 60 fps, and Marvel Rivals is proof of that. In this title, my base frame rate is acceptable, and latency isn’t a concern. In a game like this, I don’t need to resort to DLSS Super Resolution. I can just use MFG to improve the smoothness of the game with a minuscule hit to the overall system latency.
The overall latency only really becomes an issue if it’s high to begin with; as you barrel toward 0ms, even “large” spikes or dips in latency only represent a few milliseconds. If you have a higher base frame rate, even a 50% jump in latency might only move you from 20ms to 30ms. With a lower base frame rate, that same jump might move you from 100ms to 150ms. Same percentage, but the experience couldn’t be more different. Marvel Rivals is a good example of this dynamic.
However, it also exposes my concern with DLSS 4 moving forward, particularly when it comes to weaker GPUs in Nvidia’s lineup.
The flagship dilemma
Above, you can see a video of Cyberpunk 2077 with DLSS MFG running in its 4X mode. I’m not using DLSS Super Resolution here, so I’m feeding the AI model with a base frame rate of just under 30 fps — somewhere around 27 fps if I remember correctly. This isn’t how you’d want to play Cyberpunk 2077. There’s a constant motion blur over everything, along with a ton of visual artifacts.
I’m using the RTX 5090 here at 4K, and thankfully you don’t need to settle for this experience. Flick on DLSS Super Resolution to increase your base frame rate, and you’re off to the races with a smooth experience free of visual artifacts.
But we can’t all buy an RTX 5090, can we?
Something that’s important to remember about DLSS is that it’s at its best when it’s pushing the next level of visual quality; it’s at its worst when it’s making up for hardware that can’t hit the mark. DLSS Super Resolution looks and performs best at 4K. At 1080p, the performance gains are smaller and the visual quality is worse. DLSS Frame Generation is spectacular if you’re trying to climb into the triple digits when you already have a playable base frame rate. It falls apart with visual artifacts and unmanageable latency if you don’t walk in the door with that playable frame rate. The same problems are here with MFG, and they’re magnified.
Eventually, Nvidia will move onto RTX 50-series GPUs lower in its product stack. It hasn’t officially revealed its budget offerings, but I’d be shocked if we didn’t see an RTX 5060 at some point. It’s sure to be a popular GPU, likely due in no small part to DLSS 4. But will the GPU itself be powerful enough to capitalize on DLSS 4? Will MFG even be an option for players who can’t get to that 60 fps mark in demanding games?
The idea brings me back to the Surface Laptop Studio 2. When the laptop was announced with an RTX 4060, I wrote about how Nvidia is enabling a new tier of gaming experiences on hardware that otherwise wouldn’t be suited for prestige titles like Cyberpunk 2077. You may get some visual artifacts, and you may not get perfect upscaling. But it’s at least possible to play these games thanks to Nvidia’s dedication to shortcuts on the rendering pipeline.
That dynamic is thrown off when looking at MFG. It can provide what looks like a high frame rate, but without enough power to back that up, it falls apart. The video above is a clear example of that. My concern is that the bottom of DLSS 4 will fall out from under it when weaker hardware is brought into the mix. You can compromise on upscaling; you can even compromise on frame generation in titles that aren’t latency-sensitive. Here, with MFG, it’s hard to accept the compromises if you can’t feed the AI with at least 60 fps.
This isn’t a critique of Nvidia. I don’t believe a $300 graphics card — or however much the RTX 5060 ends up being — should match the experience of a $2,000 graphics card like the RTX 5090. However, it is an important note. It’s easy to look at the benchmarks and dream up what DLSS 4 could do for budget GPUs, but I’m holding my breath until those GPUs are here to see if they’re powerful enough to even take advantage of the feature.
That’s something that will need to be settled once those GPUs are here, though. For now, DLSS 4 is very impressive. If you feed it with at least 60 fps, it feels like wizardry. Add on top of that a new transformer model that promises even larger image quality improvements over time, and it’s clear that Nvidia wants DLSS to remain at the forefront of PC gaming tech.
