Let’s be fair to Nvidia first. AI upscaling is one of the best things to happen to PC gaming in the last decade. It lets mid-range graphics cards run games that would otherwise crawl, and it keeps older cards useful for years longer than they would have been.
But somewhere along the way, a clever tool became a marketing number. And the number has started to mean less than it used to.
The short version
- Upscaling (DLSS, FSR, XeSS) is a real, free performance win. Use it.
- Frame generation makes motion look smoother, but doesn’t make the game respond faster.
- Frame rates quoted “with frame generation” aren’t comparable to normal frame rates.
- AI characters are interesting, but they’re still demos more than games.
The bit that’s genuinely great
Upscaling renders a game at a lower resolution, then uses a trained model to rebuild a sharper image at your screen’s resolution. Early versions were blurry. Modern ones, starting with DLSS 2 in 2020 and improving a lot since, often look as good as native resolution, and sometimes better, because they clean up shimmering edges.
The result is simple: more frames, same picture, no extra cost. AMD (FSR) and Intel (XeSS) have their own versions, and AMD’s latest also uses machine learning. Competition has made all of them better.
If a game supports upscaling and you aren’t using it, you’re probably leaving performance on the table.
Where it gets slippery
Frame generation is different. Instead of rendering frames faster, the graphics card draws extra frames in between the real ones. DLSS 3 added one generated frame for every rendered one. DLSS 4’s multi frame generation, on RTX 50 cards, adds up to three.
On screen, that looks smoother. But the game is still only running at the original rate underneath. Your inputs are still read at that rate, and generated frames add a little delay. A game running at 30 frames per second with three generated frames in between will show “120” on the counter, and still feel like 30 when you move the mouse.
That’s why Nvidia’s launch claim that an RTX 5070 would deliver RTX 4090 performance drew so much criticism. It was only true if you counted generated frames, and those aren’t the same thing.
How to use frame generation well
Used properly, it’s a good tool:
- Start from a decent base. It works best when the game already runs at around 60 frames per second or more. Below that, the delay and visual glitches become noticeable.
- Turn on Nvidia Reflex (or AMD’s Anti-Lag) to claw back responsiveness.
- Use it for big, slower games, like flight sims, strategy and cinematic single-player games. Skip it in competitive shooters, where responsiveness matters more than smoothness.
What about AI characters?
Nvidia’s ACE tools promise characters you can talk to naturally, with voices and replies generated on the fly. We’ve seen impressive demos and a few early uses in real games, such as AI co-players and life sims with characters that make their own plans.
It’s genuinely interesting technology. But the hard part was never generating a reply. It’s writing characters that are worth talking to, keeping them consistent, and making sure they don’t say things the developer would regret. Until we see that solved in a game people play for weeks, we’ll treat it as a promise rather than a feature.
What we’d like to see
Graphics card makers should quote frame rates three ways, clearly labelled: native, upscaled, and with frame generation. All three are useful to know. Only one of them tells you how the game will feel.
That’s how we’ll report it when we test graphics cards. And when a box says a card is “four times faster”, we’ll tell you which of those numbers it means.
Where next
Planning an upgrade? Our step-by-step PC build guide covers how to pick a graphics card, and our display guide explains which refresh rate is worth chasing.
