Readers, I have spent two full days in the benchmark pits to tell you what you’ve already guessed: the GeForce RTX 5090 is very fast, too expensive, and laden with more AI tech than Philip K. Dick’s cheese dreams. At least two of those points will, I’m sure, send the average graphics card shopper running, especially at a time when even game developers are growing suspicious of generative AI and its many-thumbed, robot-voiced nostrums.
Yet while there’s not much to be done about the RTX 5090 costing at minimum £1939 / $1999, hundreds more than the infamously spenny RTX 4090, its suite of more purely performance-focused artificial intelligence tools is – dare I say it – quite neat. These range from Multi Frame Generation (MFG), which is basically DLSS 3 frame gen but up to twice as fast, to DLSS 4’s general upscaling enhancements and even the ability to apply newer DLSS versions to older games. All this will come to the rest of the RTX 50 series as well, with some trickling down to the enter RTX range, so maybe the RTX 5090 is best understood not as a practical GPU purchase in itself but as a Picadilly Square-filling advert for its more affordable siblings can do.
There’s definitely a sense of Nvidia going “Sod it” with this one, cutting the sensibility brakes to produce something that even hyperenthusiasts might find a bit too mad. Besides the price, this is by far the biggest PSU-muncher to ever wear a GeForce badge: the Founders Edition I tested demands a thundering 1000W power supply and is rated to hoover up to 575W at once, a threat that it made good on by pulling 578W in Cyberpunk 2077. In comparison, the 450W-rated RTX 4090 looks like it can be powered by potatoes. Credit to the redesigned dual-fan cooler for keeping a lid on temperatures, which I nonetheless measured peaking at a very manageable 74°c.
On that note, if there’s something the RTX 5090 does brilliantly without any AI caveats, it’s the industrial design. This might actually be – granted that it’s not a high bar – the best-looking graphics card I’ve seen in my life, a modern yet mature slice of precision-cut aluminium. Also, unlike the RTX 4090 FE, it isn’t so big that it could feasibly have come from the age when computers needed their own cupboards. It’s a solid 2cm thinner, to be exact, finally arresting and reversing the trend of ever-fattening premium GPUs.
Nvidia GeForce RTX 5090 review: 4K benchmarks
Ultra HD is the RTX 5090’s spiritual home, just as with the RTX 4090 and RTX 3090 before it. Since I had all three to hand, I introduced them to the new test rig I’d piloted with the Intel Arc B580, which combines 16GB of DDR5 RAM with an Intel Core i9-13900K.
In terms of unaltered, native-rez games performance, let alone any frame gen gubbins, the RTX 5090 is inarguably the new quickest card of them all. It’s the first ever to smash through 100fps in Cyberpunk 2077 with these settings, and provided you’ve got at least a 144Hz monitor (comparatively not a big ask, if you can afford £2K’s worth of graphics card) then you could see a visible improvement over the RTX 4090 in other tough games like Total War: Warhammer III and F1 24.
Nevertheless, it’s hard to ignore how the jump from the RTX 4090 to the RTX 5090 isn’t nearly as big or as consistent as the former’s improvement over the RTX 3090. Nvidia’s top two GPUs are almost level in Horizon Forbidden West and Assassin’s Creed Mirage, and the differences remain minor once you start adding upscaling. Quality-level DLSS in Forbidden West, for example, got it up to 137fps on the RTX 5090 and 125fps on the RTX 4090: hardly a grand leap forward.
On the other hand, the RTX 5090’s Blackwell architecture does appear a little better at handling the performance hit from ray tracing – and this was already one of Nvidia’s strong suits. After enabling Ultra-quality RT effects, Metro Exodus dropped from 163fps to 117fps on the RTX 5090 and from 135fps to 87fps on the RTX 4090: that makes for reductions of 28% and 36% respectively.
Then there’s DLSS 4, and its most numbermaking component, Multi Frame Generation. This works similarly to the frame generation that DLSS 3 introduced for the 40 series: the card takes data from the frames that the GPU is rendering and uses AI to generate lookalike frames that can be slipped in-between the rendered ones, boosting overall framerates in exchange for a touch of added input lag. However, whereas the DLSS 3 version creates up to one AI frame for every rendered frame, MSG can generate three or four, resulting in scenes like this:
That is indeed the RTX 5090 running fully pathed-traced Cyberpunk 2077 and Alan Wake II, at 4K, at – as far as your eyes are concerned – 200fps. And crucially, 4x MFG doesn’t significantly increase latency over what classic 2x frame generation already did, nor does it look any worse in terms of image quality.
That’s not to say it’s immaculate sorcery. Input lag is input lag, and while MFG doesn’t make this tradeoff noticeably worse, it also doesn’t make it any better. Cyberpunk 2077 could also get quite blurry when making rapid camera movements – an issue I didn’t really have with AW2 and Dragon Age: The Veilguard, to be fair – and it’s important to remember that no matter how many imitation frames an AI generator can add, it’s not going to make these games feel faster in your hands. That 218fps will only ever have the mouse snappiness of a 64fps game, because those frames – the ones delivered via the usual PC pipeline – are the only ones that can accurately reflect your inputs.
All that in mind, the visual smoothness is a stark upgrade, and if you’re not playing something that needs hyper-accurate twitch aiming… why not? Especially if the underlying GPU is strong enough to get comfortable ‘real’ framerates out of max-quality settings like path tracing, in which case frame generation becomes the luxury it really should be. Not a crutch that games without adequate low-end settings can use to pump their numbers up.
MFG is also but one part of DLSS 4, which covers a range of tricks that both are and aren’t exclusive to this new Nvidia generation. The RTX 5090 thus also has the duty of showcasing some tech that will benefit GPUs going right back to the RTX 20 series, such as DLSS 4’s new ‘Transformer’ model and support for ‘DLSS overrides’ in the Nvidia app.
Firstly, the Transformer model is a rejigged technique for how DLSS applies its best-in-class anti-aliasing. It runs slightly slower than the preexisting ‘Convolutional Neural Network’ model, averaging 120fps in Cyberpunk 2077 at 4K/Ultra versus the 132fps of the older version. But it does look sharper and more detailed: a lot of Nvidia’s promotional comparison images have focused on textures, but what I found most noticeable was the improvement to fine-edged objects like grass and wire fences. DLSS already looked better than its rivals FSR and XeSS on these, but the Transformer model cleans and sharpens them up even further, bringing the overall picture closer than ever to true native-rez quality. Good stuff.
Secondly, DLSS overrides could be a very big deal in bringing newer, superior versions of DLSS to games that either never updated their support or lacked it in the first place. It still requires a degree of support in the Nvidia app – you can’t just go slapping DLSS 4 on anything ever made – but there will be a batch of 75 compatible games at launch, and the options are impressively flexible. You could force a game to use DLAA anti-aliasing if it was never officially implemented, for example, or take a game that has Convolutional Neural Network DLSS support and upgrade it to the new Transformer model. And, if you’ve got a 50 series GPU like this one, you can have a game with DLSS 3 frame gen capability also work with full 4x MFG, as I did with Dragon Age: The Veilguard to get those benchmark results above.
I’m still not seeing the value in a lot of what Nvidia does with AI, up to and including their attempts at creating creepy, artless NPCs. But it’s much, much easier to get on board with these DLSS advancements, which devalue nothing, steal from nowhere, and could well breathe life into older hardware as much as they enrich the new.
Nvidia GeForce RTX 5090 review: 1440p benchmarks
Before we get too lovey-dovey, though, time for another reminder of the RTX 5090’s limitations. Buying a top-shelf GPU and hooking it up to less demanding 1440p screen might seem like galaxy brain setup-building to some – the PC equivalent of those Fiats with massive engines poking out the back because they won’t fit inside. In practice, your mismatched rigmobile will just send you crashing into a CPU bottleneck:
Again, the RTX 3090 is left well behind, but at this resolution the RTX 5090 is effectively only as powerful as the GPU it replaces. You could potentially open up the gap again by laying on thick with the ray tracing, but still, the results underwhelm: Cyberpunk 2077, with Psycho RT effects and Quality DLSS, averaged 124fps on the RTX 5090 and 111fps on the RTX 4090. All that cash for an extra 11fps? Not a great deal. MFG isn’t really useful here either, as you’re already getting so many frames shoved down your eyes that regular 2x frame generation will push supporting games beyond the point where you’ll seldom be able to perceive a load more.
Nope, 4K is where the RTX 5090 is at. But then, trying to be practical with a card like this kind of feels like missing its point. Yes, we once had the GTX 1080 Ti and the RTX 2080 Ti, GPUs that could top a range while still being somewhat attainable high-enders. I miss those days too. But is the RTX 5090 not more of an aspirational device anyway? A concept model that’s just happened to escape the design wing and sneak into the production factory?
Let’s be real: you were never going to buy this. I was never going to buy this. Even the people who might buy this, presumably when they’re not busy executing their generals, probably have at least one eye on the RTX 5080 as well. It’s just so far beyond us normies that there doesn’t seem to be much sense in hating it either. The GPU itself won’t care, and Nvidia certainly won’t. They’re probably already planning the RTX 6090 for £2,500.
Instead, let’s just see this for what it really is, an eye-grabbing exercise for DLSS features that are far more likely to make waves than any one single graphics card. You’ve got me there, RTX 5090, even if you aren’t getting my money.
This review is based on a retail unit provided by the manufacturer.