Nvidia GeForce RTX 5090 review: The fastest gaming GPU ever, but with a catch

The times they are a changing. With Moore’s Law in tatters, even Nvidia has had to admit that it can’t go on cranking out colossally more powerful gaming GPUs in every generation when the transistors haven’t become any smaller. Just as CPUs now only offer an incremental speed bump every couple of years, GPUs are going the same way in terms of raw processing power. The answer to this conundrum, according to Nvidia, is to harness the power of AI to give your gaming frame rate a helping hand, but is it enough? We put the new Nvidia GeForce RTX 5090 through its paces to find out.

We’ll cut straight to the chase on one point, which is that, whatever your views on AI, the RTX 5090 still remains the best graphics card you can buy right now in terms of performance. The RTX 4090 has been phased out, AMD has no competition in this league, and the RTX 5090 is also more powerful than the RTX 4090 in every respect anyway. In some cases the differences are disappointing when you consider that the new Nvidia flagship’s MSRP is $400 more than the RTX 4090, but Nvidia has some tricks up its sleeve in other games.

The biggest one is DLSS 4 with Multi Frame Generation, which was only available to test in a handful of games at the time of writing, but which we’ll continue to test in our benchmarks as more games start to support it. Put simply, while the RTX 4090 could generate one extra frame for each one rendered by the GPU, the RTX 5090 can generate two or even three frames. The result is transformative, as you’ll see in our results, but first let’s take a step back and look at what’s in the new GPU.

Specs

Make no mistake, the flagship Blackwell gaming GPU is a monster, and this mammoth new slab of silicon occupies a massive amount of space on the tiny RTX 5090 PCB found inside the Founders Edition. Nvidia has increased the size of the die (the actual chip, rather than just the package) from 608.5mm² on the RTX 4090 to 750mm² on the 5090, and it’s surrounded by the 16 chips that make up the card’s huge 32GB frame buffer of 28Gbps GDDR7 VRAM, up from 24GB on the 4090.

Not only that, but the VRAM is also hooked up to a super-wide 512-bit memory interface. Combine that with the speed of the GDDR7 VRAM, and you get a huge total memory bandwidth of 1,792GB/s, up from the already-huge 1,008GB/s on the 4090.

Meanwhile, the chip itself is packed to the rafters with Nvidia’s Blackwell architecture building blocks. These include 11 of Nvidia’s Graphics Processing Clusters (GPCs), which are the largest blocks of an Nvidia GPU. The GB202 Blackwell GPU on which the RTX 5090 is based has 12 GPCs (one of them is disabled in the 5090), so there’s also room for an RTX 5090 Ti in the future if Nvidia wants to make it, and can yield enough fully working chips. It’s this fully-enabled GPU that’s shown in the block diagram below.

In the meantime, though, those 11 GPCs subdivide into 170 of Nvidia’s Streaming Multiprocessor (SM) units, each of which contain 128 of Nvidia’s CUDA cores for traditional game rendering, a 4th-gen RT core for ray tracing and four of Nvidia’s 5th-gen Tensor cores for AI applications such as DLSS.

The end result is a chip with 21,760 CUDA cores, 170 RT cores, and 680 Tensor cores, which are significant increases over the figures of 16,384 CUDA cores, 128 RT cores, and 512 Tensor cores for the RTX 4090. Nvidia has also expanded the size of the L2 cache pool that’s shared by all the GPCs, from 73,728KB on the RTX 4090 to 98,304KB on the 5090.

If all these numbers are meaningless to you, the TLDR here is that Nvidia has made a bigger chip with more stuff in it than the RTX 4090. So, whatever happens with AI and DLSS, the RTX 5090 is also more powerful than the RTX 4090 in every area, with and without ray tracing, and with and without DLSS.

The downside to packing in all this gear is that, despite there being a change in GPU architecture from Ada to Blackwell, the new chip is still produced on TSMC’s 4N (technically 5nm) process, meaning the transistors are basically the same size. That’s why the 5090 chip is so much bigger than the 4090, and also why the power draw has increased from 450W to a massive 575W.

DLSS 4

By far the biggest selling point of the RTX 5090 is its support for a key aspect of Nvidia’s new DLSS 4 suite of AI technologies. This is called Multi Frame Gen, and it’s exclusive to the RTX 5000 series. Nvidia says 75 games will support Multi Frame Gen when the RTX 5090 comes out at the end of January, including Hogwarts Legacy, Stalker 2, Indiana Jones and the Great Circle, Alan Wake 2, and Diablo 4. However, we were only able to test it in a handful of games for this review.

Multi Frame Gen represents Nvidia’s aim to offer the same level of performance increase between GPU generations that we’ve seen before, but using AI rather than traditional rendering. It’s this tech that means the RTX 5070 outperforms the 4090 in Marvel Rivals, and it basically uses AI to generate three extra frames between each genuinely rendered pair. The result works surprisingly well in our experience, with barely any graphical glitches and an enormous increase in frame rate.

Nvidia says that it originally wanted to introduce Multi Frame Gen earlier, but that it had trouble with frame pacing, and the Tensor cores needed to be quick enough to calculate the optical flow correctly. This, says the company, is why it requires the new Tensor cores in the RTX 5000 cards, as they have more computational horsepower.

However, Nvidia has also made another change to DLSS with this fourth generation, which will be available to all Nvidia RTX GPUs, from the 2000 series onward. This makes a fundamental change to Nvidia’s DLSS Super Resolution tech, which displays the text, menus, and HUD elements of a game at your monitor’s native resolution but renders the 3D portion of the game at a lower resolution to improve performance, using AI to upscale it to your monitor’s native resolution.

Up until now, DLSS Super Resolution has been based on a Convolutional Neural Network (CNN) AI model, but Nvidia has now moved this to a more sophisticated transformer model, which comes with a substantial improvement to image quality. The result is, ahem, transformative, and it largely banishes many of the problems with ghosting, smearing, and blurriness that previously earned some criticism for DLSS.

We were given a beta pre-build of Cyberpunk 2077 using the new transformer model of DLSS, which can be enabled in the graphics settings, and it looks significantly better than the old CNN model. The image is sharper and there’s less ghosting too. If you look at the image we snapped from Cyberpunk 2077 above, you can see that the ‘Silk Road’ text looks a bit blurry on the CNN model, but is significantly sharper when you switch to the Transformer model.

Bear in mind that these screenshots, and the ones further down below, were all taken with the game resolution set at 1,920 x 1,080, using DLSS on the Performance setting, rather than 4K, as the lower starting resolution magnifies the defects of DLSS, enabling you to clearly see the differences.

It’s also interesting to compare the native 1080p screenshot here, as while it’s a little sharper than the new DLSS, as you can see around the red light arrows, the latter’s use of AI to smooth out the text has made it look less distorted. Basically, DLSS has gone from a feature that used to degrade image quality to one where you have to really look to see the differences. Given that DLSS is now supported in hundreds of games, this is great news for all owners of Nvidia RTX GPUs.

Likewise, if you look at the pattern on the bowls in the screenshot below, you can see that the intricate edges are much more blurred with the older CNN model of DLSS, while the transformer model looks much sharper. It’s worth right-clicking the image below and opening it in a new tab so you can properly see the difference.

In addition, a new feature called DLSS Override has also been added to the Nvidia App, which enables you to force Multi Frame Gen into games that already support frame gen, and force the new transformer model of DLSS Super Resolution as well. We tried this out with Marvel Rivals, and it worked really well.

As standard with no frame-gen enabled, the game averaged 161fps. Enable standard frame gen in the graphics menu, and it climbs up to 223fps. However, if you then force 4x Multi Frame Gen in DLSS Override, the average rockets up to 404fps, with seemingly no discernible impact on image quality. It’s genuinely impressive stuff.

Neural rendering

In addition to DLSS 4, Nvidia is also adding a number of features to the RTX portfolio under the banner of neural rendering, for which the company says it’s specifically optimized the new Blackwell architecture. Again, this makes strong use of AI and Nvidia’s Tensor cores, with the latter now becoming a part of the graphics pipeline as neural networks are placed directly in games.

As with Multi Frame Gen, the problem for Nvidia here is that these features aren’t supported in any games yet, but we did have a chance to play Half-Life 2 RTX with neural rendering at CES, and were very impressed by the impact. In particular, the use of Neural Radiance Cache made a big difference to the depth of shadows and realism on the paved floor, and the use of translucent materials in RTX Skin made headcrabs look much more realistic, for example with light shining through the skin to make it glow pink and red, rather than just looking flat.

One other key feature that’s worth mentioning here is Nvidia’s new Mega Geometry tech, which uses the RT cores in the GPU, rather than the Tensor cores, to enormously increase the number of triangles in a ray-traced scene. The company showed off this tech in its Zorah demo at CES, and the idea is that game engines can now ray-trace scenes with full geometry, rather than moving down to a low polygon count in order to improve performance.

The result is much more detail in objects, thanks to all those extra triangles, but the depth and shadows of those objects and characters are now also much more realistic. The image above shows a scene in the Zorah demo on the left, the normal triangle composition in the middle, and the triangle composition with Mega Geometry enabled on the right. As you can see, the number of triangles increases enormously when the tech is enabled in Nvidia’s Zorah demo, and the frame rate is quicker too.

The Founders Edition card

One area where we can unreservedly lavish praise on this card is the cooler design. It’s a work of engineering brilliance, especially when you consider the huge output of this GPU. To see how it works, let’s take a trip back in time to the flow-through design used on Nvidia’s GeForce RTX 3000 and 4000-series Founders Edition cards.

The idea behind these coolers was to direct cool air from the bottom of the graphics card with a fan on the bottom left, where it was then directed over the GPU and VRAM via a heatsink and heatpipe assembly, and then hot air was exhausted straight out the top, thanks to a fan on the top right.

It works really well with the average airflow system found in the best PC case designs, with the proviso that you really don’t want to install one in a case setup where there’s no room for the bottom fan to breathe.

Nvidia’s now honed that idea further, so instead of a fan on the top and bottom, there are two fans on the bottom, which direct air straight through two matching vents on the top, with the heatsink and heatpipe assembly in between. This is very different from your average third-party graphics card, which directs cool air through the front via the fans, but all the hot air is then dispersed out the back and sides.

The secret behind the new cooler’s design is to mount the new GPU on a very small PCB in the middle, so the fans genuinely do direct air straight in and out of the cooler.

There are other PCBs inside the card, for example to handle the outputs on the I/O plate, but the idea is that the cool air goes in from the fans, straight over the heatsink fins, and then straight out the top, where it can be expelled by your case’s airflow system.

The result is a card that takes up just two expansion slots, making it significantly smaller than the RTX 4090, despite being more powerful, though it is just as long and tall. In fact, the RTX 5090 is SFF ready, meaning we can expect to see it used in some of the best mini gaming PC designs in the future. Remarkably, this cooler’s efficient airflow design also means it keeps noise minimal, which is unusual for a two-slot cooler.

You can hear the fans going when the GPU is running at full load, but it’s surprisingly quiet compared to many cards we’ve tested. There are also a few slot vents cut into the edges of the card, just to make sure any hot air that doesn’t make it through the main rear vents can escape, as shown in the image above.

Finally, one extra tweak to the Founders Edition is the positioning of the 16-pin 12V power socket, which is now at an angle on the top edge, rather than sticking straight out.

This means there’s room to take your 12V cable through a routing hole on the side of your motherboard tray and straight into the socket, without making any 90-degree bends and looking very neat to boot (as we’ve shown in the image below with a CableMod cable). Let’s hope this leads to fewer problems with melting graphics cards than we saw with the RTX 4090.

How we test

To gauge the performance of the RTX 5090, we run a number of benchmarks using real games, rather than synthetic benchmarks. Each test is run three times, recorded with Nvidia FrameView, and we report the mean average of the results, discarding any obvious anomalies. We report two figures for frame rates – firstly, the average, which gives you an idea of the general frame rate you will achieve. Secondly, we report the 1% low, which is an average of the lowest one percent of results recorded during the benchmark.

This is a more reliable indicator of performance than the minimum, as it removes outliers, such as moments where a Windows system event causes the game to stutter, which is unrelated to the performance of the GPU. The 1% low is what you can expect the actual minimum frame rate to be in these games at these settings.

GPU test system specs

• CPU AMD Ryzen 7 7800X3D
• CPU cooler Corsair H100X RGB Elite
• Memory 32GB 6,000MHz G.Skill TridentZ RGB, CL28
• Motherboard MSI MAG X870E Carbon WiFi
• SSD 2TB WD Black SN850X
• PSU Corsair RM1000X Shift

Benchmarks

Cyberpunk 2077

We’re going to go straight to the most exciting test in our suite for this card first, as we were given access to a pre-build beta of Cyberpunk 2077 that allowed us to enable Multi Frame Gen. We ran the game at 4K with DLSS Super Resolution on the Quality setting, and using the new transformer model, and then enabled the highest setting of Multi Frame Gen, to see what it could do.

The results are spectacular, enabling you to play the game at 4K with full path tracing enabled at an average of 207fps, and a solid 1% low of 148fps – if you have a 144Hz+ 4K gaming monitor, then you can genuinely max out this game’s ray tracing settings on it now. Comparatively, the RTX 4090 only averages 82fps in this test, which is still perfectly playable, but in a different league in terms of frame rates, while the 5090 averages 110fps with standard frame gen enabled.

I took some time to play the game with Multi Frame Gen enabled too, and it’s remarkably smooth and responsive, while also looking fantastic with path tracing and the new, sharper DLSS Super Resolution model. If maxing out the eye candy in your games is a top priority, then Multi Frame Gen is a game changer. If you’re playing at 1440p then the results are even better, with the RTX 5090 averaging a super-smooth 337fps, compared to 148fps on the RTX 4090 with standard frame gen enabled.

We also decided to have a crack at this game without frame gen enabled, just to get an idea of the RTX 5090’s increased rendering power with ray tracing enabled. Forget path tracing, even the standard Ultra ray tracing preset is too much for the RTX 4090 at 4K without any help from DLSS. The RTX 5090 shows a definite improvement here, with the average jumping from a spotty 43fps on the 4090 to a playable 59fps on the 5090. That’s a 37.2% increase in frame rate, which is a strong improvement in ray tracing power.

Call of Duty Black Ops 6

Now we’re going to skip to a very different type of game. Despite not having any ray tracing features, Call of Duty Black Ops 6 is seriously demanding in terms of shader power at its top Extreme settings. We’ll start at 4K, where the RTX 4090 averages 120fps. That figure climbs to 146fps with the RTX 5090, which is a 21.6% performance increase. Step down to 1440p, and the difference between the cards is just 11.5%, which is disappointing for a new graphics card that costs so much money.

There’s no Multi Frame Gen support for this game right now either, but enabling DLSS Super Resolution on the Quality setting, along with standard frame gen, enabled us to run the game at 196fps on the RTX 5090. That’s a decent frame rate, but it’s not the hugely transformative jump we sometimes see with this tech, and it shows how much the RTX 5090 relies on Multi Frame Gen to deliver the big performance improvements it promises.

Indiana Jones and the Great Circle

We love this game, but the Indiana Jones and the Great Circle system requirements are absolutely brutal, with ray tracing required at every level. Step up to the full RT mode with path tracing and it’s an absolute GPU killer, with even the RTX 4090 only able to run it at 77fps with both frame gen and DLSS Super Resolution enabled.

Neither of the RTX 4080 cards can cope with this workload at all, as you can see in the graph below, as they only have 16GB of VRAM, so it will be interesting to see how the new 16GB RTX 5080 fares in this test too.

Sadly, a pre-build of the game with Multi Frame Gen wasn’t available to us for this review, and the DLSS Override options were grayed out in this game’s profile, so we were unable to force it on. The game will support Multi Frame Gen when the RTX 5090 launches, though, and we had a chance to play Indiana Jones at 5K on an RTX 5090 at CES, using Multi Frame Gen, so we know it can work. We’ll benchmark the game with Multi Frame Gen as soon as we can.

In the meantime, we’ve tested it using a custom run-through of a part of the Raiders of the Lost Ark level at the start of the game. Here, the RTX 5090’s average of 96fps and 1% low of 83fps is a decent result for this game at its maximum settings (Supreme preset, full RT), but it’s not a transformative jump over the RTX 4090, representing a frame rate increase of 24.7%.

We also ran this game at its Ultra preset without any upscaling or frame gen enabled, again just to get an idea of the new GPU’s extra rendering power without any help from AI. The RTX 5090 averages 126fps at 4K with these settings, which is a long way in front of AMD’s current top-end GPUs, but it’s only 20fps ahead of the RTX 4090, showing an improvement of 18.9%.

What this shows is that the RTX 5090 is really at its best when its new RT cores can stretch their legs with path tracing, and its Tensor cores are working their magic with upscaling and frame gen. We’re hoping that Multi Frame Gen support will provide a solid boost in this game when it’s updated later this month.

F1 24

Next up is F1 24 at Ultra settings, which includes ray tracing. We tested this game without any upscaling or frame gen tech first, where the RTX 4090 averages a solid 98fps at 4K. This improves to 137fps with the RTX 5090 installed, marking a decent 39.8% performance improvement between the two GPUs. Again, this shows that the RTX 5090 generally offers a solid increase in games with lots of ray tracing features, but less of a benefit when it comes to traditional rasterization.

There’s no Multi Frame Gen support for this game as yet, but it does support standard frame gen, as well as DLSS Super Resolution. We enabled the latter on its Quality setting, and also switched on frame gen, to see what it would do to the frame rate, and the result was a substantial improvement, with the average climbing all the way to 231fps.

However, we did see occasional glitches in this fast-paced game with frame gen enabled, which were much harder to spot in the other games we tested with frame gen. Dropping down to 1440p also showed a solid result for the RTX 5090, with the new GPU averaging 202fps, even without any help from DLSS. That’s a jump of 35.6% compared to the RTX 4090.

Doom Eternal

Finally, our Doom Eternal benchmark offers a great way to see a GPU’s raw rendering power, as it scales really well as more shaders are added. As with Call of Duty Black Ops 6, though, this older game shows the limited benefits of the RTX 5090 over the 4090 when it comes to non-ray tracing shader power.

Testing at 4K with the Ultra Nightmare settings (and no ray tracing), the RTX 5090 averages a stunning 496fps, which is obviously more than fast enough for anybody’s needs. Notably, though, the RTX 4090 isn’t very far behind at all on 440fps, marking just a 12.7% performance boost from upgrading to the 5090. It’s possible this is in part due to the game starting to become CPU limited at these settings, which seems ridiculous to say for a game running at 4K – but it still shows some of the limitations of the RTX 5090.

We then added ray tracing to the mix to see just how high the frame rates can climb with all the eye candy enabled in this game. Again, though, while the RTX 5090 is definitely faster than the 4090, the difference is minimal. The 324fps average at 4K is undoubtedly an incredible result, and again close to double that of the Radeon RX 7900 XTX. However, it’s only 11% quicker than the RTX 4090 in this test, which can churn out a fantastic average of 292fps.

Even worse, at 1440p, the performance difference between the two GPUs is just 5.6%, though this is even more likely to be a case of the CPU starting to be the limiting factor for going any faster. Thankfully, the next game in the series, Doom: The Dark Ages, will support Multi Frame Gen with DLSS 4 support when it launches, but this aging Doom title does show the limited benefits of the RTX 5090 over the 4090 if you’re not going all out on ray tracing and DLSS.

Power draw

With 21,760 CUDA cores buried inside it, and using the same 4N (5nm) process size as the RTX 4090, this massive new GPU was never going to be power-frugal, but it’s worth emphasizing just how much power it needs. Nvidia specifies that it will draw up to 575W, which means it still only requires a single 12V 16-pin power connector, but overclocked cards could end up drawing even more power.

Nvidia recommends a 1,000W PSU as a minimum for this GPU, and we agree, having tested it. On our generally power-efficient Ryzen 7 7800X3D test rig, our whole system drew a massive 682W from the mains at peak load while running F1 24 at 1440p with Ultra settings, and it may well draw more when it spikes. That’s too close for comfort for an 850W PSU, especially if you’re using a power-hungry CPU, such as the Intel Core i9 14900K.

As a point of comparison, our test rig draws 524W from the mains in this test with the RTX 4090 installed, and 422W with the RTX 4080, so the RTX 5090 is seriously demanding when it comes to power.

Price

The Nvidia GeForce RTX 5090 price is $1,999, which, let’s not beat about the bush, is a ridiculously high price for a graphics card. That’s $400 more than the RTX 4090 at launch, and third-party overclocked cards with RGB lighting are almost certainly going to be even more expensive. When Founders Editions inevitably sell out, in fact, you may well struggle to find an RTX 5090 for $1,999.

Nvidia didn’t build this graphics card to offer good value for money, but to offer the very best to those who can afford it, and it’s double the price of the next GPU down in the range, the RTX 5080.

There are a few factors at play here, including the cost of adding 32GB of fast GDDR7 VRAM, and Nvidia’s huge R&D costs for developing its new AI rendering features, but in terms of bang per buck for gamers, two grand is an enormous amount of money for what’s often a small upgrade in performance. If the RTX 5090 cost the same amount as the 4090, it would be a no-brainer, but at $400 more it’s harder to recommend.

Verdict

Nvidia has bet the house on AI with the GeForce RTX 5090, and the results aren’t as clear-cut as they’ve been with previous generational upgrades. We can argue the toss about whether using AI to generate frames and upscale resolution is cheating or not, but with the ability to cram more and more silicon into a small space diminishing, Nvidia had to come up with another way to improve performance, and the solution it offers is AI.

Does it work? In some ways yes, and in other ways no. We’re genuinely impressed by the improvement in image quality Nvidia has achieved by moving DLSS Super Resolution to the new transformer model, and Multi Frame Gen is also transformative in games that support it. In our tests, Marvel Rivals runs at frankly ridiculous frame rates, and Cyberpunk 2077 is genuinely playable at 4K with full path tracing enabled. With 75 games supporting the tech at launch, the RTX 5090 looks set to massively improve frame rates in several key titles.

However, there’s no getting around the fact that 75 is a drop in the ocean in the PC’s enormous catalog of games, many of which don’t even support ray tracing, let alone Multi Frame Gen – the biggest title of 2023, Baldur’s Gate 3, is a case in point. DLSS may indeed be supported by hundreds of games now, but Nvidia will need to work hard to get Multi Frame Gen implemented in as many games as possible if it wants the RTX 5000 series to take off.

The same also goes for the new neural rendering features, which clearly have the potential to make a big difference to gaming graphics, but need game support. In some ways, the situation is reminiscent of when Nvidia first implemented hardware T&L with the GeForce 256, or shaders with the GeForce 3. In both cases, barely any games supported the new tech at launch, but later became firm standards.

The difference here is that those GPUs also at least offered a decent step up in basic rendering performance over their predecessors, and the same can’t be said of the RTX 5090. We’ll be eagerly watching the spread of Multi Frame Gen, and benchmarking it as it appears in new games, which may well result in us revising this review in the future.

In the meantime, the RTX 5090 is only really worth considering if you want the very best GPU and you’re prepared to pay the extremely high price for it. For all its reliance on AI to get the most out of it, it’s still the fastest gaming GPU you can buy, and the competition simply can’t get anywhere near it. If you already own an RTX 4090, then we advise holding off before upgrading, but it’s a decent step up from an RTX 3000-series GPU.

In short, the RTX 5090 has the potential to be a killer graphics card, but it isn’t there yet. With its massive price of $1,999, the performance difference over its two-year-old predecessor is also largely disappointing, especially when it costs $400 more.

Multi Frame Gen is good, but it’s not yet widely supported enough for a $1,999 graphics card to get a sky-high review score for it. If and when Multi Frame Gen and neural rendering take off, the RTX 5090 may turn out to be the must-have graphics card for those with the cash to spare, but there’s little need for most RTX 4090 owners to upgrade just yet.