Every couple of years, Nvidia launches an extremely expensive, extremely powerful graphics card that brings PC gaming into a new generation. That is what the Nvidia GeForce RTX 5090 ultimately is, but the way it brings next-generation performance is unconventional, to say the least. Because in a lot of games, the performance uplift over the RTX 4090 isn’t quite as steep as you’d expect – at least when DLSS Frame Generation is taken out of consideration. With the next generation of Nvidia’s DLSS for both upscaling and frame generation, however, we get leaps in image quality and performance that feel even greater than what we see with a typical graphics generation.
How much of an upgrade the Nvidia RTX 5090 is going to be for you, then, is ultimately going to depend on the games you play, the resolution you play those games at, and whether you’re ok with an AI algorithm generating extra frames. For a lot of people playing games on anything less than a 4K monitor with a 240Hz refresh rate, this upgrade is simply not going to make a lot of sense. But if you do have a high-end display, these AI-generated frames are going to feel like a taste of the future.
RTX 5090 – Specs and Features
The Nvidia GeForce RTX 5090 is built on Blackwell, Nvidia’s high-end architecture that’s already powering the data centers and supercomputers behind many of the most popular AI models. That should give you an idea of what the RTX 5090 is especially good at, but Nvidia didn’t neglect the, well, non-AI parts of the card.
With the 5090, Nvidia found a way to shove more Streaming Multiprocessors (SMs) into the same amount of GPCs (Graphics Processing Clusters), which means more CUDA cores – 21,760, up from 16,384 in the RTX 4090. That makes up for a 32% uplift in the amount of shader cores over the previous generation, and is where a bulk of the raw gaming performance comes from.
Each SM also has four Tensor Cores and one RT Core, just like its predecessor. That means you get 680 Tensor Cores and 170 RT cores, compared to 512 and 128, respectively, for the RTX 4090. The 5th-generation Tensor Cores are tailor-made to boost AI performance, with this generation adding support for FP4 operations, which should make AI workloads less dependent on VRAM.
All of this silicon is coupled with 32GB of GDDR7 VRAM, or video memory. This is a generational shift from the GDDR6X memory in the RTX 4090, and should be faster and more power efficient than the previous generation. But because the RTX 5090 requires a staggering 575W of power, a huge increase over the already power hungry 4090, power efficiency isn’t exactly Nvidia’s main goal with this graphics card.
Because the new Tensor Cores are more efficient, Nvidia shifted the entire DLSS algorithm to run on a Transformer Neural Network (TNN), rather than a Convolutional Neural Network (CNN). This shift won’t necessarily improve your framerate when you enable DLSS, but Nvidia claims it will improve image quality, and mitigate issues like ghosting and other unwanted artifacts.
Nvidia did more than just make an under-the-hood change to the way DLSS works. Team Green also introduces Multi-Frame Generation, which takes the Frame Gen tech introduced with the RTX 4090, makes it more efficient and smooth, and allows it to generate multiple frames off of each rendered image. This drastically improves frame rate, but should probably only be enabled if you’re already getting a decent frame rate, just like the last generation version.
The Founders Edition
The Nvidia GeForce RTX 5090 requires 575W of power, which is much more than the 450W of the RTX 4090. More power inherently means more heat, which means increasingly powerful cooling solutions are needed. Looking back at the RTX 4090 and even the RTX 3090, the Founders Editions were these giant, triple-slot graphics cards that took up a ton of room, and straight up wouldn’t fit in some PC cases.
Before I saw the RTX 5090, I was expecting something even bigger and more unwieldy. However, somehow, it’s smaller. Nvidia was able to make a 575W graphics card fit in a dual-slot chassis with a dual fan configuration. And it works.
Throughout my time with the RTX 5090, which included both my standard testing suite and playing games with DLSS 4 enabled to test multi frame generation, the temperature maxed out at around 86°C, even while its power consumption peaked at 578W. That’s a high temperature, to be sure, and higher than the RTX 4090’s 80°C, but it’s not high enough to throttle, and that’s all that matters.
Nvidia was able to do this by shrinking down the PCB (printed circuit board) to a little square and placing it in the middle of the graphics card. The two fans are placed on each side of where the PCB is, with a heatsink that runs through the width of the card. These fans then take air in through the bottom of the card and shoot it straight through the top of the card and out through your exhaust fans in your PC case. In fact, this graphics card doesn’t even have exhaust vents under the output ports at the rear of the card, unlike previous generation designs.
But it’s immediately apparent that the RTX 5090 Founders Edition follows a similar design language to the last couple of generations. The center of the card has the same silver ‘X’ design as the RTX 4090, with a gunmetal-gray chassis surrounding the black heatsinks. On the outer edge of the graphics card, you get a ‘GeForce RTX’ logo that lights up with white LEDs, too.
Next to that logo, you’re going to find the power connector. While it looks very similar to the 12VHPWR connector of the last generation, it’s actually a new 12V-2×6 power connector. The difference is minor, but it’s supposedly more efficient than the last-generation connector. Maybe Nvidia will be available to avoid controversy around its power connectors melting with this generation – only time will tell.
Nvidia does include an 12V-2×6 adapter in the box, which takes four 8-pin PCIe power connectors in order to provide the required 575W of power. But what’s nice is that the connector on the graphics card itself is now angled, and facing the back of the graphics card, which should make connecting the cable much easier. The power connector seems more secure this time around, too.
The hidden benefit of this design is its ability to be slotted into smaller PC builds. You don’t need a giant case to run the RTX 5090, like you did with the RTX 4090 and 3090. However, it’s very likely that third-party designs from the likes of Asus and MSI are going to be much larger than Nvidia’s Founders Edition.
DLSS 4: Fake Frames?
When Nvidia revealed the RTX 5090, it claimed that it could boost performance by as much as 8x. The actual number isn’t that high, but the RTX 5090 can deliver extremely high frame rates in the most demanding games, but not exactly through traditional rendering. Because while the RTX 5090 does deliver a decent increase in raw rasterization performance, the real next-generation benefit is in its ability to generate extra frames to increase your frame rate.
DLSS 4 introduces ‘Multi-Frame Generation’, a next-gen version of the Frame Generation introduced with DLSS 3 and the RTX 4090. But it’s more than just using the same method to just produce more frames. And the secret behind it is a new AI Management Processor, or AMP core in the RTX 5090 – along with other RTX 5000 graphics cards. The AMP allows the graphics card to essentially assign work to different parts of the GPU, something that was traditionally handled by your CPU. But because it’s physically located on the GPU, it’s able to do this much more efficiently.
According to Nvidia, the AMP and the 5th-generation Tensor Cores allowed it to create a new frame generation model that’s both 40% faster than the original frame generation model, while requiring 30% less memory. This new model only needs to run once on each rendered frame, which then can create 3 AI frames. Something like this would naturally introduce latency, but Nvidia found a way around that, too.
The AMP runs a Flip Metering algorithm, which paces out the frames in order to reduce input lag. Nvidia claims this is why multi-frame generation won’t work on RTX 4000 graphics cards, as the last-generation frame generation relied on the CPU for frame pacing, which would introduce much more latency than the new model, which runs entirely on the GPU itself.
To be clear, this isn’t a magic button to get good performance. Just like the previous generation, you only really want to enable this if you’re already getting a passable frame rate. If you’re not already getting around 60 fps with Frame Gen disabled, turning it on can introduce significant latency problems. That’s why it pairs best with DLSS upscaling also enabled, in order to maximise your performance.
When the RTX 5090 hits store shelves on January 30, DLSS 4 will work in a wide array of PC games that already support DLSS 3 Frame Generation. However, while working on this review, I only had access to two games with this technology enabled, and both were on beta builds – Cyberpunk 2077 and Star Wars Outlaws.
And I was surprised how well it worked. In Cyberpunk 2077 at 4K, on the Ray Tracing Overdrive Preset, with DLSS on Performance mode, the RTX 5090 gets 94 fps. That’s not bad for a game with full ray tracing. When I turned on DLSS 2x frame gen – the same as is supported by the RTX 4090 – that framerate increased to 162 fps. That’s a 2x improvement over just plain ol’ DLSS. However, when I cranked the frame generation to 4x, that’s 3 AI frames per rendered frame, that number went all the way up to 286 fps – more than my 4K display can actually render.
It’s a similar story with Star Wars Outlaws. Playing at 4K with all the settings cranked up to max, I was able to get up to around 300 fps with DLSS 4 enabled – and that’s up from about 120 fps without frame generation.
When I saw these framerates, I was straight-up expecting to see artifacts and weird spikes of lag. However, I only really saw one broken texture in Star Wars Outlaws, and it’s something I wouldn’t have noticed if I wasn’t actively looking for problems. It’s hard to believe, but Multi-Frame Generation actually does work, you just need to have an extremely high-end 4K display to benefit from it, at least with the RTX 5090.
It’s easy to write off this performance as ‘Fake Frames’, and you wouldn’t necessarily be wrong, especially because you need good baseline performance to make it a good experience. But it is going to be genuinely useful for anyone with a high-refresh, high-resolution display. It’s also important to keep in mind that I was only able to test it in a handful of games. Nvidia claims that 75 games will support DLSS 4 when the RTX 5090 hits store shelves on January 30, and there’s a decent possibility that it won’t work flawlessly in at least one of those games. For the time being, though, it looks like it works extremely well.
RTX 5090 – Performance
The Nvidia GeForce RTX 5090 is an incredibly powerful graphics card, but testing this thing was a journey. In 3DMark, the RTX 5090 proved itself to provide a generational improvement over the RTX 4090 in terms of raw performance. However, things get a lot more complicated when I test actual games. In the vast majority of games, the RTX 5090 is bottlenecked by my CPU, even at 4K – and I paired it with the Ryzen 7 9800X3D, the fastest gaming processor on the market right now. For most people who already have a high-end graphics card, upgrading to this $1,999 GPU isn’t going to make a world of difference – the games just aren’t there yet. This is a graphics card you buy to set yourself up for future PC games, like The Witcher 4.
I want to note that I did not enable DLSS 4 for any of these comparative benchmarks, and everything was tested on the public drivers available at the time. That means all non-5090 Nvidia cards were tested on Driver version 566.36, and all AMD cards were tested on AMD Adrenalin 24.12.1. All games were tested on their latest public builds, too.
In 3DMark, the RTX 5090 is up to 42% faster than the RTX 4090. In the Speed Way benchmark, the RTX 5090 scores 14,399 points to the RTX 4090’s 10,130, making for a 42% performance uplift. Similarly in Port Royal, the ray tracing test, the RTX 5090 scores 36,946 points to the 4090’s 25,997 points, which is also a 42% performance leap. What’s more impressive is how far Team Green has come since the RTX 3090. That graphics card got 5,619 points in Speed Way and 13,738 points in Port Royal, meaning anyone that skipped the last generation can get a 2.5x performance jump. That’s just 3DMark, though, and not necessarily reflective of real-world gaming performance.
In Call of Duty Black Ops 6, however, we start to see the big issue with the RTX 5090 in today’s games – a severe CPU bottleneck. At 4K Extreme settings, with DLSS set to ‘performance’, the RTX 5090 gets 161 fps, compared to 146 fps from the RTX 4090. That’s just a 10% performance difference, and definitely not what I’d call a ‘next-generation’ performance increase. However, when looking back at the RTX 3090, the 4-year-old graphics card gets 91 fps, which means a nearly 2 x performance increase.
What’s wild is that the RTX 5090 even shows signs of CPU bottleneck in Cyberpunk 2077, one of the most demanding PC games on the market right now. At 4K, with the Ray Tracing Ultra preset and DLSS set to performance, the RTX 5090 gets 125 fps, compared to 112 fps from the RTX 4090 with the same settings. Similarly to Black Ops 6, this is just a 10% performance increase. That scaling just gets worse at lower resolutions, though, with the RTX 5090 getting 153 fps at 1440p and 156 fps at 1080p.
I test Metro Exodus: Enhanced Edition with DLSS disabled, because it’s the only upscaling solution in that game, and I need to get an honest comparison with AMD cards. This makes it one of the most demanding tests in my suite, and gives the RTX 5090 a chance to stretch its legs a bit. At 4K with the Extreme preset, the RTX 5090 gets 95 fps, compared to 76 fps from the RTX 4090 and 44 fps from the Radeon RX 7900 XTX. But even without the upscaling, the RTX 5090 only gets a 25% improvement over the RTX 4090, even if it more than doubles the 39 fps from the RTX 3090.
Red Dead Redemption 2 is getting up there in years, but it’s still a gorgeous game. At 4K with every setting maxed out, and DLSS set to performance mode, the RTX 5090 gets 167 fps, compared to 151 from the RTX 4090 and 92 fps from the RTX 3090. That means in this game, which admittedly doesn’t even use ray tracing, the RTX 5090 gets a paltry 6% performance uplift over the RTX 4090.
Total War: Warhammer 3 is an interesting test these days, because it doesn’t support ray tracing or upscaling, and gives a clear picture on raw rasterization performance. The RTX 5090 impresses here, delivering 147 fps to the RTX 4090’s 107 fps. That’s a 35% performance uplift and close to the potential performance difference demonstrated in 3DMark. However, it’s still a far cry from the 67% performance difference enjoyed by the RTX 4090 over the 3090.
Assassins Creed Mirage is a weird one. For some reason, when I first benchmarked this game, the RTX 5090 was getting terrible performance. Its game clock was limited to 772MHz, and was giving me around 50 fps at 4K. I was able to get around that problem, but even when it was resolved the 5090 only got 172 fps, which is lower than the RTX 4090 at 183 fps. What’s worse, is that the framerate was extremely spikey with microstutters. That’s bad, obviously, but it’s very likely that this is a driver bug, and as such should be treated as an outlier.
Black Myth: Wukong, like Cyberpunk 2077, is an extremely demanding game that will push any GPU to its limits. The RTX 5090, however, averaged 104 fps at 4K, with the Cinematic Preset and DLSS set to 40%. The RTX 4090, with identical settings, got 84 fps. That’s a 20% uplift in favor of the RTX 5090.
In Forza Horizon 5, the RTX 5090 averaged 216 fps, compared to 210 fps from the RTX 4090, which is essentially in the margin of error. This is an aging game, to be sure, but the CPU bottleneck means there’s essentially no difference between these two cards at this resolution.
Nvidia really wants us to believe that Moore’s Law is dead and GPUs that deliver a giant uplift over their previous-generation counterparts are going to grow more rare over time. I don’t know if that’s true – I’m not an engineer – but regardless, in most games, the RTX 5090 doesn’t exactly deliver next-generation performance over the RTX 4090 – at least not to the extent that the latter card thoroughly trounced the 3090 back in 2022.
That’s not to say the RTX 5090 is a bad graphics card. No matter how you slice it, the RTX 5090 is now the fastest graphics card on the consumer market, and that’s not nothing. The problem is that a lot of games can’t really take advantage of the extra power offered by the Blackwell GPU. That’s something that will absolutely change over time, but it also means there’s little reason for someone with an RTX 4090 to upgrade to the new hotness.
Instead, the Nvidia GeForce RTX 5090 is betting its existence on the future of AI-powered gaming. DLSS 4 uses AI to greatly increase frame rates, which is definitely a sight to behold. This graphics card is therefore best for gamers that want to be on the cutting edge, and are willing to bet $1,999 (at least) on an AI gaming future. For everyone else, the RTX 4090 is going to be more than powerful enough for the next few years.
Jackie Thomas is the Hardware and Buying Guides Editor at IGN and the PC components queen. You can follow her @Jackiecobra
