Nvidia GeForce RTX 5090 Founders Edition

PCMag editors select and review products independently. If you buy through affiliate links, we may earn commissions, which help support our testing.

Nvidia has been the rock of the graphics card industry for years, but its position in the market has never felt more cemented in place than now. AMD has always provided Nvidia with feisty Radeon competition, but the high end of the market has been Team Green territory for a while. And when Nvidia announced the $1,999 GeForce RTX 5090, no one doubted that this would be the world’s fastest consumer graphics card yet. And so it is: The Nvidia GeForce RTX 5090 has an astonishingly huge graphics chip under its shell, at a similarly colossal price. While you could argue that Nvidia's pricing and design indicate the RTX 5090 is not primarily for gaming PCs (and really, it's more of a content-creator card and an AI powerhouse), its ability to deliver an unrivaled high-fidelity gaming experience earns it our Editors' Choice award in the ultra-elite category of graphics cards. That said, for most mere mortals, the GeForce RTX 5080, which is launching in parallel on Jan. 30, will be the more realistic and fitting choice in a high-end, cutting-edge GPU.

'Blackwell' and the Nvidia GeForce RTX 5090

The Nvidia GeForce RTX 50 series features a new graphics architecture called “Blackwell,” which Nvidia says has instructions-per-clock (IPC) improvements over the "Ada Lovelace" architecture used in the last-generation RTX 40 series. Blackwell also has improved ray-tracing and AI capabilities and a host of feature updates, but the community doesn't yet have a fixed idea of exactly how much impact these changes will have.

This is mainly because Nvidia has also gone big with the new GeForce RTX 50-series GPUs. The Nvidia GeForce RTX 5090’s GPU die, known as GB202, has more than 92 billion transistors and roughly 33% more shaders than the Nvidia GeForce RTX 4090. The RTX 5090 also has a wider 512-bit memory interface compared with the RTX 4090’s 384-bit interface, which connects to 32GB of GDDR7 RAM. That bank of video memory is higher-capacity and faster than the 24GB of GDDR6X in the RTX 4090.

These significant increases in resource counts and the larger and faster RAM pool would substantially increase performance even if the RTX 5090 used the same architecture as the RTX 4090. That’s partially why it’s hard to differentiate the benefits the underlying architecture brings. Understanding Blackwell’s advantages is also tricky because many involve AI applied in new ways.

AI seems like almost all Nvidia ever talks about nowadays, and you'll find some immediate advantages from the RTX 5090's AI hardware. However, large companies like Amazon, Meta, and Microsoft will be the ones who primarily enjoy most of those advantages, stuffing as many RTX 5090 cards as they can into server racks. AI at the training level just isn’t the tool of everyday people quite yet. Using a new UL benchmark, we can at least test the performance of the AI hardware to some degree for local processing of some popular large language models (LLMs), and in the right situations, you can also take advantage of Nvidia’s DLSS technology in games, which also runs on the AI hardware.

With the RTX 50 series, Nvidia released DLSS 4 as the newest version of its DLSS technology. This is essentially a turbo-boosted version of its DLSS 3 technology focused on frame generation. You'll find two key differences between DLSS 4 and DLSS 3. First, it can generate more frames and introduces a multiplier setting that seems relatively straightforward, letting you get roughly two, three, or even four times the number of frames (with the extras AI-generated) at the flip of a switch. The other change is that DLSS 4 introduces a new "Transformer" AI model that is supposed to produce better image quality than DLSS 3’s "Convolutional," or CNN, model at the cost of a few frames.

You'll find several other gaming-related AI features that Nvidia introduced with Blackwell, like RTX Neural Shaders, but for now, those features aren’t something I can test. The preceding link gets into some of those new technologies; before we can test these features, they will need to be adopted by game developers, and that simply hasn’t happened quite yet. Nvidia stated during CES 2025 that cooperation with Microsoft is in the works to incorporate support for these features into DirectX. So we may not have to wait long before we see the first games to support RTX Neural Shaders and its ilk.

Design: A Look at the Nvidia GeForce RTX 5090 Founders Edition

The star of today’s review is the Nvidia GeForce RTX 5090 Founders Edition. Nvidia’s board partners will have their version of this card, and a select few, like MSI's GeForce RTX 5090 Suprim Liquid SOC we briefly previewed, will liquid-cool the RTX 5090. This is probably a wise decision for a graphics card that can pull 575 watts by itself, which is more than most PCs use. However, Nvidia stuck with a cleverly designed air cooler for the Founders Edition, not a million miles from what it's done on earlier Founders Edition cards.

The bulk of what you see when looking at a graphics card is its thermal solution, and again, given the power draw, it’s surprising to see the RTX 5090 Founders Edition is considerably smaller than the last-gen RTX 4090 Founders Edition. Both cards have identical length and width measurements of 11.97 and 5.39 inches, respectively, but the RTX 4090 is a triple-slot card that is 2.4 inches thick, while the RTX 5090 is a dual-slot card at 1.57 inches thick.

Nvidia was able to do this with the RTX 5090 Founders Edition by compacting the circuit board into the middle of the graphics card. If you tore it down, you'd find more than one circuit board inside the RTX 5090 comprising the internals of the card to maximize the use of space. Doing this enabled Nvidia to extend heat pipes toward the left and right sides of the card and add fans that blow air straight through the card on both sides to get the most cooling performance in the available space. (See also our unboxing and first hands-on tour of the RTX 5090.)

Technically, liquid cooling can do an even better job here, but that requires mounting a radiator somewhere in your case that connects to the GPU via liquid-filled tubes, which takes up more space overall. It also complicates installing the graphics card, so Nvidia had understandable reasons to stick with the air-cooled design.

Like all of the Nvidia GeForce RTX 40-series GPUs, all 50-series RTX cards employ the Nvidia-specific 12VHPWR power connector to connect the graphics card directly to the power supply. Some key changes here, though, are welcome improvements. On the RTX 5090, the cable connector is now angled at 45 degrees rather than straight up out of the card, which reduces the chance of pressure on the connection, harsh cable bending, or possible shearing on the fragile connection. Also, the bundled four-socket power connector that runs between your four PSU leads and the card connector uses much softer, more flexible wiring for less chance of fatal bends and easier cable routing.

The Nvidia-supplied adapter that comes with the RTX 5090 converts four eight-pin PCIe power connectors to a 12VHPWR, as most power supplies still don’t natively support this connector. You'll want at least a 1,000-watt power supply to play it safe with this card; in any case, you'll need that kind of mega-wattage, or close to it, to get four 8-pin connectors to start with. This GPU also has a single HDMI port and three DisplayPort connections.

Testing the GeForce RTX 5090 Founders Edition: Setup and Competition

We are transitioning to a new graphics card test bed and have yet to finish retesting our extensive library of previously reviewed graphics cards. The new graphics card testbed is based on a 16-core AMD Ryzen 9 9950X processor set on a Gigabyte X870E Aorus Master motherboard with two 16GB sticks of DDR5 RAM. This RAM supports AMD EXPO memory profiles and was set with the EXPO profile to operate at 6,000MHz.

We use a large triple-fan 360mm water cooler to manage CPU temperatures. We opted to use two PCIe 4.0 SSDs (a 2TB boot drive and a 4TB secondary drive) for the storage, one for Windows 11 and everything that isn’t a game, and the second exclusively to hold games. The system also has a 1,500W Corsair power supply, which gives us adequate overhead given how power-hungry modern graphics cards are becoming.

Given the Nvidia GeForce RTX 5090’s enormous resources, $2,000 price tag, and 575W power draw, it should be no surprise that this card has little to no competition. AMD simply doesn’t bother to compete in this market segment with its consumer cards, as the audience shrinks considerably at this price. In many ways, the RTX 5090 competes better with workstation graphics cards, and the people likely to buy an RTX 5090 will likely consider one of those, too.

The only cards worth mentioning as potential competitors to the RTX 5090 in the consumer-card field are the Nvidia GeForce RTX 4090, the Nvidia GeForce RTX 5080, and the AMD Radeon RX 7900 XTX. None competes in performance with the RTX 5090, though the RTX 4090 does come closest. Instead, these cards stand out as compelling alternatives that are generally more affordable—if you consider a $1,000 graphics card (for the cards other than the 4090) affordable.

Testing the GeForce RTX 5090 Founders Edition: Synthetic Tests

The test scores we recorded in 3DMark show what you would expect from the RTX 5090 specs and price. It was notably faster in all of the 3DMark tests. The RTX 4090 got a little close to it in the 3DMark Solar Bay ray-tracing test, but given we didn’t see that anywhere else, I think the RTX 5090’s score could have been limited by our AMD Ryzen 9 9950X from going even higher (despite it being one of AMD's top-most chips right now).

We also ran the GPU-maker-specific 3DMark screen optimization test to test DLSS performance on Nvidia cards and FSR performance on the AMD cards. Interestingly, the AMD Radeon RX 7900 XTX was faster with DLSS and FSR off in this test, but these tests aren’t quite the same when measuring FSR (here, FSR 2.0) versus DLSS (DLSS 3 with frame generation, as tested) performance, so 1:1 comparisons aren't fair.

We only list these tests together as they are both from 3DMark and tailored to test broadly similar technologies from each company. Though these tests are similar, they aren’t entirely the same, just as DLSS and FSR aren’t entirely the same, so it's best not to dwell too closely on comparing the AMD and Nvidia numbers in this specific trial. It does show the RTX 5090 still leading the Nvidia pack, though, and ahead of the AMD cards, with DLSS or FSR, respectively, on.

Testing the GeForce RTX 5090 Founders Edition: Content Creation Tests

The content creation benchmarks we used for test-driving the RTX 5090 rely on a mixture of CPU and GPU performance. We started with the Adobe Premiere Pro 24 video editing app, tested via a canned sequence of video rendering and editing tasks in Puget Systems' PugetBench for Creators utility. The Nvidia GeForce RTX 5090 once again came out on top, but the Nvidia GeForce RTX 5080 wasn’t that far behind, and it suggests the RTX 5080 could be a better buy for this sort of work. While the RTX 5090 likely would have held a more significant lead if the GPU was used more heavily in this test, this mixed-usage type of scenario is likely more representative of genuine real-world use, which bears keeping in mind.

If you do a lot of content creation work in the 3D modeling app Blender, whether you should buy an RTX 5090 is much more straightforward. In all three Blender tests, the RTX 5090 was leagues faster than the tested pack here. The same is true for the Chaos V-Ray 6 benchmark, a rendering-application test derived from the V-Ray 3D graphics plug-in, used in industries like movie and game production, pro design, and architecture. (Note: The CUDA-core performance in this test shows the RTX 4090 ahead of the 50-series cards, but Nvidia tells us this is a quirk with the RTX 50 series and Chaos V-Ray 6 CUDA benchmark that hasn’t been fixed yet.)

Testing the GeForce RTX 5090 Founders Edition: UL Procyon AI Tests

Testing AI performance is difficult, particularly as AI is intended for a wide range of tasks--training versus inferencing, for one--and is impossible to test in an all-encompassing way. We have no perfect answer to testing AI, but we can at least measure the AI performance of graphics cards using UL’s Procyon AI Text Generation test for now...

This test presses the GPU for text-creation inferencing tasks via four popular AI large language models (LLMs) and produces an overall score in each case. (It uses versions of Mistral, Microsoft's PHI, and two flavors of Meta AI's Llama.) It also reports how many tokens (or discrete units of text, in the context of text generation) the hardware can produce per second with each model. It also measures the time it takes to produce the first token, as getting as much work done as fast as possible is basically AI’s mission statement.

In these tests, the Nvidia GeForce RTX 5090 once again came out on top across the board. The lead that the RTX 5090 held over the RTX 4090 and the RTX 5080 wasn’t quite as large as we might have expected, but high-end AI hardware is often deployed in server environs with multiple GPUs, so cumulatively, the RTX 5090's lead could add up to a bigger advantage in such a scenario.

Testing the GeForce RTX 5090 Founders Edition: DLSS, FSR, and XeSS Testing

In reconfiguring our GPU testbed for 2025 and beyond and recasting our graphics tests, we are reducing how many DLSS, FSR, and XeSS tests we conduct. All three technologies are helpful, but comparing their effectiveness is complicated. Though similar, DLSS, FSR, and XeSS don’t all work quite the same way and have different impacts on image quality and latency. This makes comparing them head-to-head somewhat misleading, as it's hard to quantify or accurately weigh off those image-quality and latency tradeoffs against frame-rate boosts and a perception of greater smoothness. The options for adjusting DLSS, FSR, and XeSS settings make them even more complicated to test in an apples-to-apples way. This is only compounded by DLSS 4's multi-frame frame-generation technology and alternative AI models.

We will continue to test DLSS, FSR, and XeSS, particularly with the standalone Black Myth: Wukong benchmark, which mandates the use of one of these technologies, but while comparing the test scores here, you should keep in mind that the image quality isn’t necessarily the same even though we use the same graphics presets.

In Black Myth: Wukong, the Nvidia GeForce RTX 5090 was the fastest card tested at all resolutions with and without frame generation enabled. None of the other cards came close, and its 1% lows—a new measure we are reporting, the average of a benchmark run's lowest 1% frames—were better overall, too. The AMD Radeon RX 7900 XTX showed better 1% lows with its version of frame generation enabled than most of the Nvidia cards, but the RTX 5090 still held a safe lead.

Testing the GeForce RTX 5090 Founders Edition: Ray-Tracing Game Tests

Shocking no one, the RTX 5090 continued to outpace its competition in PC games that support ray tracing. This lead was significant enough that the RTX 5090's 4K performance was often more in line with the 1440p speeds from all of the other cards in this group.

The only test in which others in this test lot came close to the RTX 5090 was Call of Duty: Modern Warfare III, which appears to have been bottlenecked by our AMD Ryzen 9 9950X. You won't find many faster desktop processors than this one, though, so it’s not an issue that is easy to overcome.

We could belabor the point, noting that the RTX 5090 had a performance lead of between 118% and 176% over the AMD Radeon RX 7900 XTX (AMD's top consumer entry) in Cyberpunk 2077 or its performance in other areas, but these tests paint a simple picture of the RTX 5090's total domination.

Testing the GeForce RTX 5090 Founders Edition: Raster Game Tests

The CPU bottleneck mentioned in Call of Duty once again appeared when we tested Total War: Three Kingdoms. All of the tested GPUs came in relatively close to each other at the 1080p resolution here. The RTX 5090 also hit that limit at 1440p, but none of the other graphics cards did. At 4K, it performed almost twice as fast as the AMD Radeon RX 7900 XTX or the Nvidia GeForce RTX 4080.

We also appear to be hitting that limit with the Nvidia GeForce RTX 5090 in Shadow of the Tomb Raider. Fewer cards hit the limit this time, but the RTX 5080 matched the RTX 5090 at 1080p due to the processor limitations.

Testing the GeForce RTX 5090 Founders Edition: Power and Thermal Tests

We use a Kill-A-Watt wall meter to measure the power consumption of the host system as a whole; swapping out the card changes the relative overall consumption on each trial. Though the Nvidia GeForce RTX 5090 can pull more power than any graphics card we have previously tested, as you can see in the chart, this doesn’t happen to the same degree in every application. It hit a lower power peak than the AMD Radeon RX 7900 XTX in our Adobe Premiere Pro test. It did use more in Blender, but its bigger delta of performance in that test makes up, at least in part, for the increased power draw.

In games, we saw just how power-hungry the RTX 5090 can be. It pushed the system to draw up to 859W in Cyberpunk 2077 and used just a little less than that in F1 2024. Its performance numbers again argue that this could be worth the trade-off, but from these numbers overall, it looks like the Nvidia GeForce RTX 5080 will be the more energy-efficient of these new cards, if that concerns you.

If you buy an Nvidia GeForce RTX 5090 and put it in a well-ventilated case (our testbed is a big, roomy Asus ProArt PA602 tower), you shouldn’t have too much to worry about keeping it cool, as you can see in the relative thermal measures above. The card stayed relatively chill throughout our testing process. However, you will want to ensure your PC case is well-ventilated, as eventually, the heat could build up in it and become an issue.

Verdict: Who Needs This Much Power?

No doubt: The Nvidia GeForce RTX 5090 is the world’s single fastest consumer graphics card of the moment. The bigger question is, "Who is this card for?"

Though undeniably powerful, it probably shouldn’t be the top pick on your list for your next graphics card if all you're thinking about is PC gaming. It's prohibitively expensive, and while it runs everything exceptionally well, so do the other cards we tested. They don't match the RTX 5090 on paper, but anyone saying you can’t have a very enjoyable gaming experience with, say, an Nvidia GeForce RTX 4080 or an AMD Radeon RX 7900 XTX isn’t being honest with you. The RTX 5080 looks similarly compelling for that crowd. The exception would be those fixated on high-refresh-rate 4K play; this card pushes the upper limit of what is possible there.

That’s not to say we wouldn’t recommend the RTX 5090 in any situation, though. If you have an unlimited budget for your gaming PC, this is your card: Get one. You won’t have to worry about upgrading for quite a while, and it's hard to say how long it will be before the RTX 5090 isn’t good enough anymore for high-end, high-frame-rate gaming—more than just a few years, to say the least.

The RTX 5090 would also be a good pick if you do a lot of GPU-acceleration-dependent content-creation work or can leverage it for its AI abilities. (More on that latter aspect as the nascent AI-benchmarking field takes shape.) But for anyone not doing one of those things and on a more constrained budget, you could realistically build a full gaming PC with an Nvidia GeForce RTX 5080 for what the RTX 5090 costs. Consider that before buying.

Francisco La Hoz contributed testing work to this review.