Nvidia RTX 5090 Founders Edition Review

It’s that time again: Nvidia is here with its latest GPU architecture, aimed at powering the games and experiences we have to even greater levels. Revealed back at CES 2025, the RTX 5090 Founder’s Edition is Nvidia’s new flagship GPU, boasting some impressive specs while also doubling down on the AI workloads that have come to define gaming for the past six years ever since Nvidia introduced machine-learning based upscaling with DLSS during the 20-series launch. How does the RTX 5090 Founder’s Edition stack up, though? Here is our review.

Specifications

• MSRP: $1999
• Architecture: Blackwell
• Nvidia CUDA Cores: 21760
• Tensor Cores: 5th Generation; 3352 AI TOPS
• Ray Tracing Cores: 4th Generation; 318 TFLOPS
• Boost Clock: 2.41 GHz
• Base Clock: 2.01 GHz
• VRAM: 32GB GDDR7
• Memory Interface: 512-bit
• Memory Bandwidth: 1792 GB/s
• Total Board Power: 575W
• Required PSU: 1000 Watts
• I/O: 3 x DisplayPort 2.1b; 1 x HDMI 2.1b
• Display support: 4K @ 480Hz or 8K @ 120Hz with DSC
• Connector: 600 W PCIe Gen 5 cable (4x PCIe 8-pin cable adapter provided)
• Size: 2-slot form factor; 304mm x 137mm (LxW)

Unboxing and First Thoughts

When I first started to hear rumors and leaks that the RTX 5090 would need close to 600W of power to bring the GPU to life, I started to get worried I would need to go out and buy yet another case to fit the GPU into. The RTX 4090 and its subsequent GPU family are known for being chonky behemoths, with the RTX 4090 and RTX 4080 both boasting large, over-sized coolers to keep the GPU cool under load. 

Yet, when Nvidia finally revealed the RTX 5090, I was incredibly surprised to see just how small it was by comparison to the RTX 4090. The card itself is relatively thin, so thin that when I first held our review sample in my hands, the line by Genie in Aladdin came to mind: “Supreme cosmic power, itty bitty living space.” 

The card itself features a new Double Flow Through design, aimed at giving the most optimal airflow through the GPU. This was accomplished in large part because the team was able to effectively shrink the PCB of the GPU, allowing for the Double Flow Through design. To help with this cooling the RTX 50-series will utilize a 3D Vapor Chamber, as well as liquid metal to help keep the Blackwell GPUs cool under pressure.

I do like the fact that even with this redesigned cooler, the RTX 5090 still hearkens back to the visual design language that Nvidia has adopted going back to the 30-series. The Mobius strip is still predominant on the card, providing some real elegant lines across the front and back of the GPU, though it might not be as pronounced as it was on, say, the RTX 4090.

One part I do particularly love is the packaging. The RTX 40-series Founders Edition cards had some of the largest boxes I’ve ever seen. I actually had to buy a whole second shelf just to store our review samples for future testing. I was preparing to do so with this review as well, imagining that even more power would equate to an even larger box.

Pictured: Nvidia RTX 5090 (Top), Nvidia RTX 4080 Super (Bottom) for size reference

Thankfully, because the RTX 5090 and its family of GPUs are much smaller than the Ada Lovelace cards, the packaging is too. It’s also 100% plastic-free, instead made entirely of paper fiber. It’s refreshing to see box size come back down to a realistic size, and the fact that the new packaging can also be propped up as a small display on a shelf is an added touch I like as someone whose office is adorned with tech.

Despite its smaller overall size footprint, the RTX 5090 does require a fair bit of power. Compared to last generation’s RTX 4090, which clocked in at a Total Graphics Power (TGP) of 450 watts, the RTX 5090 requires a whopping 575 watts. This necessitates a powerful power supply, with a 1000-watt unit being a minimum for the GPU. 

The RTX 5090 is built using a 4NP process and boasts around 92 billion transistors on the chip. This is coupled with fast 32GB of GDDR7 VRAM, 21,760 CUDA cores, as well as Nvidia’s 5th Generation Tensor AI cores and 4th Generation Ray Tracing (RT) Cores. 

Bringing AI Back To GeForce

One of the more memorable quotes on stage at Nvidia’s CES keynote was when CEO Jensen Huang stated that AI was the house that Geforce built, and that AI was now “coming home” to Geforce. AI workloads have been a crucial part of Nvidia’s strategy for years, and we’ve seen it play out in games most notably thanks to Nvidia’s machine learning-powered Deep Learning Super Sampling, or DLSS, image reconstruction solution. This allowed games to push visuals such as real time raytracing while providing stellar image quality and, many times, better performance than if using a traditional rendering method.

This was built upon during the Ada Lovelace generation with DLSS 3’s Frame Generation. By generating every other frame, Nvidia’s solution was able to dramatically improve the performance of games, all while pushing the graphics boundaries with solutions such as full path tracing and ray reconstruction. 

Now, with the Blackwell architecture, AI is even more integrated into what Nvidia is trying to do, not only from a sheer performance level, but also when it comes to helping game developers create more beautiful, performant games. 

Chris broke down all the major details in our CES Editor’s Day post last week, which I definitely encourage you to read before making any purchasing decision. The long and short of it is that in addition to changes to how DLSS operates, swapping the old Convolution Neural Network (CNN) model to a more performant, more advanced Transformer model (TNN), Nvidia’s is introducing new features such as RTX Neural Shaders, RTX Neural Texture Compression, RTX Neural Materials, and more. 

Put together, these technologies can dramatically increase the visual fidelity of any scene with less performance impact thanks to how Blackwell handles these AI tasks. RTX Neural Materials, for example, takes that shader code and compresses it in a way that allows the GPU to process it faster than traditional rendering, while RTX Neural Texture Compression will compress and store texture data, only to recall it when needed much more efficiently than traditional texture compression models. 

All of these technologies point to a future where game visuals can, theoretically, be taken to a whole new level of real-time rendering. During the CES keynote, the Zorah demo highlighted this future, with its stunning lighting, incredibly detailed texture and complicated geometry in scenes, as well as features aimed at, to use their words,  “crossing the uncanny valley” with character models thanks to features like RTX Skin, RTX Hair and RTX Neural Faces. It’s exciting stuff.

What’s more exciting to me is the prospect of DLSS 4 and the improvements brought by its new Transformer AI model used to reconstruct an image, and of course, the performance benefits of DLSS’s Multi-frame generation. We’ve gone on record here calling DLSS itself “Black Magic,” and I still stand by that, as it’s one of the first features I look to turn on in any game now, especially since oftentimes DLSS’ image resolves into a clearer, more detailed result than even native raster rendering in many cases.

The new Transformer model replaces the initial CNN model DLSS has been using since its inception with the RTX 20-series six years ago. The Transformer model allows image quality with technologies like Mutli-frame Generation to stay high while reducing ghosting across an image, resolve finer details such as fence links or wires across a screen, and smooth out edges even more for an overall cleaner image.

DLSS’s Multi-frame Generation is a dramatic leap over the current frame generation model introduced by the Ada Lovelace GPUs last generation. Instead of generating an additional frame for each traditionally rendered frame, Multi-frame Generation can create three out of every four frames, using improvements to the Blackwell architecture to ensure for even and smooth frame pacing thanks to hardware-based flip metering. At launch, up to 75 games and apps will feature DLSS 4 support, including games like Alan Wake 2, Cyberpunk 2077 and more. 

RTX 5090 Founders Edition Benchmarks

All that said, what really matters is how the card performs under pressure. We put the RTX 5090 Founders Edition through its paces in a variety of gaming and synthetic benchmarks, aimed at pushing the card to its limits, as well as testing the new feature set of the card. While some technologies like the Neural Shaders and such weren’t available yet, we did get to try out the new DLSS 4 Transformer AI model, as well as the new Multi-frame generation in a variety of games, including Cyberpunk 2077, Marvel Rivals and Dragon Age: The Veilguard.

As part of our bench, we included comparisons to last generation’s flagship, the RTX 4090 Founder’s Edition, as well as the RTX 4080 Super Founders Edition. For AMD cards, we included the two top-line RDNA 3 cards, the RX 7900 XTX and RX 7900 XT for comparison.

For the hardware powering the tests, we put the RTX 5090 Founders Edition through the tests on our Starforge Systems Voyager II test bench, which we detailed in a separate article this year. Here are those specs for quick reference:

• CPU: Intel Core i7-14700K
• CPU Cooler: Starforge Custom Bitspower 360mm Liquid Cooler
• RAM: Teamgroup Delta RGB 32GB DRR5 6000 CL38 (2x16GB)
• Motherboard: MSI Z790 Tomahawk Wifi DDR5
• Cables: CableMod Pro ModMesh Sleeved Cable Extensions (Black)
• Primary Storage: 2TB PCIe 4.0 NVMe SSD (min 7000 read)
• Power Supply: Corsair HX1200i 1200 Watt Platinum
• RGB Fans: 6 x Bitspower 120mm ARGB
• Case: Lian Li O11D EVO RGB Mid-Tower (SMITE Custom Printing Version)

We ran through a number of titles on various game engines, focusing on games that also have ray tracing as part of the feature set. Every game in our full testing suite was run using the in-game benchmark tool to provide the most consistent run each time.When it came to testing the DLSS 4 and Mutli-frame Generation titles, these were run in a mix of in-game benchmarks and a repeatable circuit, such as walking around the Defense of the Dark Arts tower in Hogwarts Legacy or testing a replay of the same Marvel Rivals match from the same perspective each time. 

We’re changing up our approach a little bit as well this time around. For the vast majority of users, technologies such as DLSS or AMD’s FidelityFX Super Resolution (FSR) will be turned on when available, especially when playing with ray tracing features enabled. We feel it can clutter a graph by providing granular data such as performance without “X” feature turned on, then that same run with part of the feature enabled, etc. Instead, we’re testing these cards as they will be played, with the full feature set. This includes testing with features such as frame generation when applicable.

This is also a 4K card, aiming directly at high-refresh 4K gaming. We still included 1440p numbers for comparison’s sake (and since we know many people are still rocking 1440p displays), but this card is targeting 4K gaming, full stop. So our write up will focus more on those numbers more than anything else as we feel they are the most relevant.

When testing with DLSS or FSR, we used the Performance preset, and in every game we tested at the highest possible settings for either standard raster tests and ray tracing tests. We also ensured that Resizeable BAR was enabled to get the most out of the performance the test bench.

RTX 5090 Founders Edition Synthetic Benchmarks

In Firestrike Extreme, which is the 4K variant of 3DMark’s DX11 benchmark, we see the RTX 5090 Founders Edition on top in the overall score, sitting 29% higher than RTX 4090 Founders Edition, while the gap widens with the RTX 4080 Super (78% higher). Compared to AMD’s RDNA 3 cards, the gaps are wider, with the top-end RX 7900 XTX’s score improved on by 63%, while the RX 7900 XT score is beaten by almost double. 

Meanwhile, Time Spy Ultra, the DX12 test, is a little better for AMD, closing the gap a smidge. But the RTX 5090 Founders Edition is still the clear winner in every respect, sitting a nice 69% faster than the RX 7900 XT. 

RTX 5090 Founders Edition flexes mightily as well in 3DMark’s Port Royal ray tracing benchmark, thanks to tis 4th generation RT cores. Compared to the RTX 4090 FE, the RTS 5090 FE is 30% faster, while the new Blackwell flagship is 83% faster than the RTX 4080 Super. Meanwhile, the RTX 5090 FE beats the RX 7900 XTX and RX 7900 XT by 109% and 142%, respectively.

RTX 5090 Founders Edition Gaming Benchmarks

When it comes to standard rasterized gaming, the RTX 5090 Founders Edition is by far and away the best option at 4K, as shown by our testing. In Cyberpunk 2077, the RTX 5090 beats out the RTX 4090 at 4K, seeing a 47% uplift over last gen’s top card, while it’s 112 FPS at 4K is more than double the RTX 4080 Super’s 55 FPS. 

Gains are a bit more marginal in Final Fantasy XIV Dawntrail’s benchmark, with only a 3% uplift over RTX 4090, though still holding a sizable improvement over AMD’s RDNA 3 cards. Black Myth Wukong also sees the RTX 5090 hold an edge over RTX 4090, 188FPS versus 143FPS at 4K, a 31% increase in performance for 25% more money. Meanwhile, the RTX 5090 sees better 1% lows at 4K over nearly every card, highlighting not simply high framerates overall, but smoother feeling performance in motion, which is crucial in an action game like Wukong that requires fast and precise reaction times.

With ray tracing enabled, we really see the RTX 5090 unleashed, with faster overall performance than any card in our suite thanks in large part to those 4th Gen RT Cores. Even in games where both AMD and Nvidia frame generation techniques are present, the RTX 5090 still holds an edge, such as Cyberpunk 2077’s 180% uplift at 4K compared to the RX 7900 XTX, AMD’s best card from last generation. 

More stark is our testing of Black Myth Wukong, though we do think there might be something hampering AMD performance on the driver level as frame generation didn’t seem to be working no matter how many times we re-ran the benchmark (as well as did a fresh driver install on our AMD tests). At 4K, the RTX 5090 Founders Edition is 29% faster than RTX 4090 FE, while it’s 68% faster than RTX 4080 Super. 

DLSS 3 Frame Generation vs DLSS 4 Multi-frame Generation

One of the major features with the Blackwell GPUs is obviously frame generation. As part of this review process, we were able to test a preview version of Cyberpunk 2077, Alan Wake 2, Hogwarts Legacy and Star Wars: Outlaws, all with native Multi-frame Generation support, while Marvel Rivals and Dragon Age: The Veilguard had the option to force this at the driver level thanks to the Nvidia App’s Override feature.

All of these tests were run using the new Transformer AI model when able to toggle and, as before, DLSS Performance was used as the preset. 

Naturally, with more generated frames, performance should, in theory, scale even higher upwards. In  Cyberpunk 2077, DLSS 3.5 Frame Generation’s still impressive 157FPS with all the bells and whistles enabled is improved upon by 80% thanks to MFG 4x, while we see similar improvements in Hogwarts Legacy and even larger boosts in Dragon Age: The Veilguard. 

But it’s the 1% Low improvements that impress most, seeing improvements across the board. While with DLSS 3.5 the RTX 5090 Founders Edition still experienced sub-100FPS 1% Lows, in every test using DLSS 4 MFG 4x these 1% Low figures jumped to above 100FPS and, in all but Marvel Rivals we see these 1% Lows eclipse 150FPS. 

RTX 5090 Founders Edition Thermals And Power Consumption

The RTX 5090 is a power hungry card, and this was shown in our testing. Alan Wake 2 actually hit the rated 575W of TGP when pushed to full tilt, including Nvidia’s new RTX Neural Geometry, though on average, in our testing, power consumption stayed below 500W.

Thermal performance was fairly impressive as wel, with our RTX 5090 FE hitting a max temp of 76 degrees Celsius during that same Alan Wake 2 test, though on average the thermals sat around the mid-60s degrees Celsius. That’s hotter than last generation’s RTX 4090 Founders Edition ever got in our testing  (it, by comparison, never crossed the 70 degrees Celsius mark), it’s still not too worrying, especially if your card is in a large case with proper airflow.

RTX 5090 Founders Edition Final Thoughts

So what does this all mean? The RTX 5090 Founders Edition is the fastest card I’ve ever tested, for one, but its package of DLSS features and the potential that its future feature suite from RTX Neural Shaders and more offer make it a powerful package. 

I’ll just be blunt: this is a halo product, and it’s priced as such. For many people, $2000 is out of reach, and that’s okay. The exciting aspect of the RTX 5090 at the top of the stack is the potential the architecture has to scale as we start testing the rest of the GPUs in the Blackwell family. If this is the ceiling, and we can infer that each subsequent card will be an improvement on the previous card at its price point, it’s interesting to think about.

In practice, DLSS 4’s new AI model and most especially Multi-frame Generation feel transformative. The difference between the CNN and the Transformer model might be subtle, especially if you’re not entirely sure what to look for, but in motion during my time playing games such as Dragon Age: The Veilguard and Cyberpunk 2077 with it enabled I noticed less aliasing and more detail on textures across the scene. 

One of my favorite tests is to look at puddles and the reflections they show in games like Control and Cyberpunk 2077. The added clarity of these reflections might be easy to miss in motion most times while you’re focused on the action, but when they are more in your face it can help draw me in even more to the action.

Additionally, the extra detail resolved in textures just makes things look even better across a screen, whether it be the stitching on a character’s backpack or finer detail in the stonework on a nearby ruin. 

What’s better is that this isn’t a feature that is simply packaged with the 50-series cards, but every RTX GPU will be able to take advantage of the improved Transformer DLSS model. Additionally, even if a game doesn’t natively support it, you can force it through the Nvidia Override feature in the Nvidia App, which allows more customization to play the way you want. 

The part that is most exciting to me when I think about it is that since this is all powered by machine learning, this is the worst it’s ever going to be, which is saying something since it’s already impressive. It’s only (in theory) going to get better as the AI model is trained on even more data over time.

However, it’s the new Multi-frame Generation where the improvements brought by Blackwell will be felt. MFG’s ability to let you tweak the amount of frames you want the AI to generate is a nice touch, especially for those who don’t have a 240Hz 4K monitor, so the 4x MFG feature might be wasted on them. 

This level of granularity helps dial in the perfect visuals and the perfect feeling experience, tailored to your set up. I found myself playing mostly with 2x or 3x MFG depending on the game simply because I didn’t need to get 200+ in most games, especially since my main 4K gaming monitor is capped at 144Hz anyway. 

Naturally, generating these extra frames brings the concern of added latency into the equation. Yes, there is a latency hit in every game we tested, but it’s not much and, I daresay for most people, it’ll be imperceptible, especially if you play on a controller which has its own latency to overcome. At most in nearly all the games I noticed only about a 3-5ms difference between FG 2x, which is effectively what DLSS 3 has been this whole time, and MFG 4x.

The only one that was really noticeable was Alan Wake 2, where I saw a nearly 25ms jump in latency when going from MFG 2x to 4x. However, I do think that this was partly because I was playing so many benchmark tests back to back and so it was just readily noticeable. Once I was playing through my benchmark circuit after a few seconds that latency was imperceptible once again.

The game I thought it would be most felt was Marvel Rivals, thanks to it being a competitive hero shooter. Yet Rivals felt more responsive than ever. I played through about 10 matches as a variety of characters, from Doctor Strange or Cloak and Dagger and each time it just felt spectacular (and, I daresay the MVPs I scored could be down to just more responsiveness). 

The 1% Low numbers also showcase a relatively smooth experience overall. In every test in the RTX 5090 Founders Edition’s MFG 4x tests, the 1% Lows were well above 100FPS. Alan Wake 2 was the only game where I noticed some stutter, but otherwise everything was pretty overall in my testing.

In the end, the RTX 5090 Founders Edition feels like a true flagship GPU, showcasing the potential of Nvidia’s Blackwell GPU fully unleashed. It’s not cheap, but if you’re a gamer who wants the absolute best out of your rig, this is the card to buy, especially if money is no object. 

It’s an uplift in performance over the Ada Lovelace heavy hitter, the RTX 4090, with many of our tests seeing 30% improvements in performance for only 25% more money. While we don’t know how this will stack up to AMD’s upcoming RDNA 4 line of GPUs, we do know that Radeon is effectively ceding the high-end market to Nvidia at this point, meaning the RTX 5090 Founders Edition will likely sit in a class of its own, much like its Ada predecessor did. 

All in all, the RTX 5090 Founders Edition, while expensive, is incredibly impressive. From outrageously fast performance at 4K with all of the ray tracing bells and whistles turned on to its exciting suite of features aimed at making games look and run better well into the future, it’s a compelling option if you’ve the means to drop on the card. It’s power hungry, sure. But that power brings incredible performance and eye-popping visuals, which will only get better and more performant as DLSS 4 matures. 

Full Disclosure: The product discussed was provided by the manufacturer for the purpose of  review.