Nvidia RTX 5090 review: fast, but not nearly twice as fast

The RTX 5090 is a hard GPU to review. By the numbers, it’s undoubtedly the best graphics card you can buy. That’s what happens when you’re the only one in town making this class of GPU, and as it stands now, Nvidia is. If you want the best of the best and don’t mind spending $2,000 to get it, you don’t need to read the rest of this review — though, I’d certainly appreciate if you did.

No, the RTX 5090 is about everything else that RTX 50-series GPUs represent. It delivers that flagship gaming performance, but it also ushers in an entirely new architecture, DLSS 4, and the era of neural rendering. And on those points, the dissection of the RTX 5090 is far more nuanced.

Nvidia RTX 5090 specs

The RTX 5090 is angled toward PC gamers who want the best of the best — regardless of the price — but it’s also the first taste we’ve gotten of Nvidia’s new Blackwell architecture in desktops. The big change is neural rendering. With RTX 50-series GPUs, Nvidia is introducing neural shaders — along with DirectX — though we won’t see the fruits of that labor play out for quite some time.

For immediate satisfaction, Nvidia has DLSS 4. This feature is coming to all RTX graphics cards, replacing the convolutional neural network (CNN) that DLSS previously used with a new transformer model. Nvidia says this leads to a quality boost across the board. For the RTX 5090, the more important addition is DLSS Multi-Frame Generation, which promises up to 4X frame generation in 75 games on day one. DLSS 4 is coming to all RTX graphics cards, but DLSS Multi-Frame Generation is exclusive to RTX 50-series GPUs, including the RTX 5090.

Although it might seem like Nvidia could just flip a switch and enable DLSS Multi-Frame Generation on all of its GPUs, that’s not exactly the case. Nvidia says with 4X frame generation and Ray Reconstruction enabled, there are five AI models running on your GPU for each rendered frame. To manage all of that, the RTX 5090 includes an AI management processor, or AMP, which handles scheduling of these different workloads across the ray tracing, Tensor, and CUDA cores.

Outside of AI hardware, the RTX 5090 brings 32GB of GDDR7 memory. Nvidia bumped up the capacity from 24GB on the RTX 4090, though that doesn’t have a ton of applications in games. The extra memory here really helps AI workloads, where training large models can easily saturate 32GB of memory. The bigger boost is GDDR7, which is twice as efficient as GDDR6 while providing twice as high of a data rate.

Nvidia also redesigned its ray tracing and Tensor cores for Blackwell, both of which it says are built for the new Mega Geometry feature. The bigger standout for me is the media encoding engine, however. Nvidia now supports 4:2:2 video encoding, along with DisplayPort 2.1 output. Those are some significant upgrades over the RTX 4090, regardless of what the benchmarks say.

4K gaming performance

Twice as fast as the RTX 4090? Not quite. Based on my results, the RTX 5090 is about 30% faster than the RTX 4090 when the new DLSS Multi-Frame Generation feature isn’t brought into the mix. And it’s a feature you might want to leave out of the mix in some titles, as I’ll dig into later in this review. That sounds like a solid generational jump, but I went back to my RTX 4090 review for a sanity check. It’s not nearly as big as what we’ve seen previously.

With the RTX 4090, Nvidia provided over an 80% generational improvement, which is massive. Here, it’s actually more of a lateral move. The RTX 5090 is 30% faster than the RTX 4090, but it’s also 25% more expensive, at least at list price. That said, good luck finding an RTX 4090 in stock at $2,000, much less at list price. The RTX 5090 may not be the generational improvement I expected, but the reality for buyers is still that it’s the best option for flagship performance.

The average is brought down by a handful of games where the RTX 5090 doesn’t show a huge increase. In Assassin’s Creed Mirage, for example, there’s about a 17% uplift. Similarly, in Forza Motorsport, the improvement shrinks to just 14%. Those aren’t exactly the margins I was hoping for when Nvidia announced a new flagship GPU, and especially one that comes in at a significantly higher price.

Make no mistake; there are still big wins. As you can see above, I measured a massive 54% improvement in Cyberpunk 2077, which is really impressive. In the previous generation, the RTX 4090 was the only GPU that could run this game at 4K Ultra without upscaling and still achieve 60 frames per second (fps). Now, the RTX 5090 is comfortably reaching into the triple digits. This is the kind of improvement I expected to see across the board.

Cyberpunk 2077 isn’t a one-off thankfully. Although the improvements aren’t quite as large across the board, I saw similarly impressive uplifts in Horizon Zero Dawn Remastered, Returnal, and Dying Light 2. The improvement may not be above 80% like we saw in the previous generation, but there’s still a clear improvement. If you want the best of the best, Nvidia is claiming that throne with the RTX 5090.

It’s just the expectations that are important. Despite some big wins, I suspect most games will look like Black Myth: Wukong, Red Dead Redemption 2, and Call of Duty Modern Warfare 2. You’re getting a nice chunk of extra performance, no doubt, but that lift doesn’t fundamentally change the gameplay experience in quite the same way that the RTX 4090 did.

Looking over my 4K data, it became clear that the RTX 5090 establishes somewhat of a new normal. The RTX 4090 had an outsized generational improvement, as Nvidia continued to navigate the waters of how it wanted to market its flagships moving forward. The RTX 5090 is disappointing by comparison, and I’m not sure there’s much reason for RTX 4090 owners to run out and buy Nvidia’s latest. But for those that want the best, it’s hard arguing with the numbers the RTX 5090 puts up.

It’s easy to argue, however, with Nvidia’s misleading claims. We’re nowhere near twice the performance of an RTX 4090, and the company confirmed to me that it’s seeing a 30% average uplift internally, as well. That’s the kind of improvement I’d expect to see out of an 80-class card, but it looks like the death of Moore’s Law has to hit everyone at some point.

1440p gaming performance

Even at 1440p, it’s very easy to run into a CPU bottleneck with the RTX 5090. You can see that just from looking at the averages above; the RTX 5090 shrinks down to just a 22% lead over the RTX 4090. All of my data here is fresh, and run with a top-of-the-line Ryzen 9 9950X. In short, if you plan to use the RTX 5090 at 1440p, you’re giving up a serious chunk of its performance potential, and you’re probably better off with a used RTX 4090.

Forza Motorsport and especially Red Dead Redemption 2 show the problem here. The RTX 5090 is still able to squeeze out a win across games at 1440p, but the margins are much thinner. That’s not a critique of the graphics card, but it is the reality of trying to run this monstrous GPU at any resolution below 4K.

There are still some solid wins for Nvidia’s latest, particularly in games that scale well on your CPU. Cyberpunk 2077 is once again a standout victory, but you can see similarly large improvements in Dying Light 2 and Returnal. 

One game that’s worth zooming in on is Black Myth: Wukong. This is the only game in my test suite that I run with upscaling enabled by default, and it shows what can happen when forcing upscaling on at a lower resolution. The RTX 5090 is providing a 20% improvement, but as you continue to push down the internal resolution, that lead will continue to flatline.

Regardless, the RTX 5090 really isn’t built for 1440p. You can use it at this resolution, but you’re giving up a chunk of what the RTX 5090 is truly capable of.

1080p gaming performance

The idea of using an RTX 5090 at 1080p is a little silly, but I still ran the card through all of the games I tested at this resolution. Here, the CPU bottleneck becomes more extreme, pushing the RTX 5090 down to just a 15% lead over the RTX 4090. You could see that as disappointing, but frankly, I see this resolution as unrealistic for a $2,000 graphics card.

However, looking at 1080p data is still valuable, at least at a high level. It’s important to remember that DLSS Super Resolution renders the game at a lower internal resolution, so the advantage of the RTX 5090 slips a bit with DLSS upscaling turned on. The RTX 5090 can easily make up that gap with DLSS Multi-Frame Generation — and even push much further — but these results are a good reminder of bottlenecks you can run into when using flagship hardware with upscaling.

Ray tracing

Nvidia dominates when it comes to ray tracing, so it’s no surprise that the RTX 5090 enjoys a top slot among the games I tested. However, the improvements aren’t as large as I expected. Nvidia has “solved,” for lack of a better word, real-time ray tracing. Games that aren’t pushing full-on path tracing are seeing less of an improvement, largely due to the fact that lighter forms of ray tracing are fair game for GPUs as weak as the Intel Arc B580.

Dying Light 2 is a good example of this dynamic. When this game released a few years back, it was one of the most demanding titles you could play on PC. But even at 4K with the highest graphics preset and no help from upscaling, the RTX 5090 makes Dying Light 2 look like child’s play with a comfortable 90 fps average.

In Returnal, the situation is even more extreme. This is one of those lighter ray tracing games available, and sure enough, the RTX 5090 crosses triple digits without breaking a sweat, even at 4K.

Things get interesting when looking at those more demanding ray tracing games, though. Cyberpunk 2077, once again, serves as a mile marker for the RTX 5090. It’s the first GPU to get close to 60 fps at 4K with the RT Ultra preset, which is quite the achievement. Of course, it’s possible to push the RT Overdrive preset, as well — more on that in the next section — but looking at raw performance, Nvidia is pushing to new heights.

The next frontier is path tracing, and for that, I used Black Myth: Wukong. The RTX 5090 provides a great experience, even at the Cinematic graphics preset in the game. But games like Black Myth — such as Alan Wake 2 and Cyberpunk 2077 — that have a path tracing mode still need to resort to upscaling, introducing the CPU more into the mix and limiting the performance uplift. Maybe in the next few generations we’ll see a native 60 fps in this title from an Nvidia flagship.

There really isn’t much to talk about when it comes to ray tracing on the RTX 5090, and that’s exactly how Nvidia wants it. In the vast majority of games, you’re looking at rasterized performance that comfortably clears 60 fps at native 4K and can easily climb into the triple digits. Ray tracing still forces some upscaling wizardry in titles like Black Myth: Wukong, but for the most part, you can flip on ray tracing without a second thought. That’s the way it should be.

A closer look at DLSS 4

The chart above is the story Nvidia wants to tell about DLSS 4. Nvidia didn’t make this chart, nor did it tell me to make it, but there’s a clear narrative that emerges from the data here. Even factoring in PC latency, which is the main issue with frame generation technology, DLSS 4 is doing some magical things. You’re going from an unplayable frame rate to something that can fully saturate a 4K, 240Hz monitor like the MSI MPG 321URX. And you’re doing so with around half of the average PC latency as native rendering.

The devil is in the details here, however, and Nvidia has a few little devils to contend with.

Here’s a different side of the story. Above, you can see a short section of gameplay — I’m going for five stars here — in Cyberpunk 2077 with the RT Overdrive preset. I’m using DLAA for a little boost to image quality, and I’m using the 4X frame generation mode. Given that I’m playing on a 4K display with a 138Hz refresh rate, these seem like ideal settings for my setup. Watch the video, and you tell me if it looks like an ideal experience.

I can point out a lot of problems here, picking out single frames with various visual artifacts between each swipe of the mouse, but you don’t need to pixel peep to see the issue. There’s an unnatural motion blur over everything, and the edges of objects are mere suggestions rather than being locked in place. You don’t need a trained eye to see that this is a bad experience. You don’t need a point of comparison, even. You can watch this video in a vacuum and see that DLSS 4 has some clear limitations. That’s not a damning critique of DLSS 4. It’s a wonderful tool, but you need to use it correctly.

Like any frame generation tech, your experience will rapidly deteriorate when you feed the frame generation algorithm with a low base frame rate like I did in Cyberpunk 2077. Nvidia wants you to use Super Resolution to get to a playable base frame rate of near 60 fps, and then click on Multi-Frame Generation to saturate a high refresh rate display. Using Multi-Frame Generation alone, especially if you’re hovering around 30 fps, will give you a bad experience.

Cyberpunk 2077 shows the worst of what DLSS 4 has to offer, but Marvel Rivals shows the best. This is one of various games that uses Nvidia’s new DLSS Override feature, allowing you to add up to 4X Multi-Frame Generation to games with DLSS Frame Generation through the Nvidia app. Not only is the base frame rate high enough here — well over 60 fps, even with DLAA turned on — but you also have a third-person camera. There are some minor artifacts, but nothing that ruins the experience and nothing you’d even notice during gameplay.

Similarly, the artifacting isn’t nearly as bad in Alan Wake 2 as it is in Cyberpunk 2077. Here, once again, I’m starting with a base frame rate of around 30 fps and using Multi-Frame Generation to make up the difference. There are some artifacts, and I’d recommend using a combination of Super Resolution and Frame Generation instead. But the experience is at least better compared to Cyberpunk 2077 due to the camera angle.

You don’t want to just crank DLSS 4 to 4X mode and call it a day. It needs to be fed with a base frame rate of ideally 60 fps. Although the latency doesn’t significantly increase up to three generated frames — something that Nvidia should be applauded for on its own — the number of visual artifacts does. Realistically, I suspect DLSS 4 will more often run in 2X or 3X mode alongside Super Resolution. That, in a lot of games, will provide a much better experience than relying on Multi-Frame Generation alone.

Over the past few generations, Nvidia has increasingly relied on DLSS to market its graphics cards, and that same playbook is at work here. It’s just not the same selling point that it once was. Super Resolution is still pulling a lot of the weight, and even a single generated frame is enough to saturate most gaming monitors, even as refresh rates climb. There’s still a use for 4X Multi-Frame Generation, and with the right circumstances, it works extremely well. But when it comes time to spend $2,000 on a graphics card, I would seriously consider how much DLSS Multi-Frame Generation is offering over a $7 utility like Lossless Scaling.

For my money, it isn’t providing much of an advantage.

This is where you need to carefully consider your setup. You want to be using Multi-Frame Generation alongside Super Resolution in those prestige games like Cyberpunk 2077 and Alan Wake 2, and if you don’t have a monitor capable of producing that high of a refresh rate, that second or third generated frame goes to waste. Unlike DLSS 3, Multi-Frame Generation isn’t a feature that just works on its own; it needs to work as part of the rest of your gaming rig.

Great — for those in the market

Nvidia’s CEO hit the nail on the head when defending the price of the RTX 5090: “When someone would like to have the best, they just go for the best.” If there’s one thing I can say with absolute certainly, especially considering the lack of flagship competition from AMD, it’s that the RTX 5090 is the best. It doesn’t matter if it’s $1,500, $2,000, or $2,500 — Nvidia’s CEO is right when he says that the appetite for this type of product doesn’t factor in price nearly as much as more inexpensive options.

The question isn’t if the RTX 5090 is the best; it is. The question is if you need the best, and there’s a bit more discussion there. The generational improvements are here, but they don’t touch what we saw with the RTX 4090. DLSS 4 is incredible, but it falls apart when it’s not fed with the right information. And 32GB of GDDR7 memory is welcome, but it’s only delivering a benefit in AI workloads, not in games.

If you’re sitting on an RTX 4090, there’s not much reason to upgrade here. There’s a performance boost, but the real value lies in DLSS 4, and that’s something that’s very easy to get around without spending $2,000. The RTX 5090 really shines for everyone else.

Maybe you had to skip the RTX 40-series due to poor availability, or maybe the RTX 2080 Ti you have just isn’t providing the grunt that it used to. In those situations, the RTX 5090 is great. But if you’re in the market to spend $2,000 on a graphics card, you probably don’t need me to convince you.