Nvidia RTX 5090

Subscribe to our newsletter here!

* Required field

E-mail *

Name *

Nvidia's Blackwell platform is, unsurprisingly, packed with exciting new technologies and ways of doing things. We're looking at the Nvidia RTX 5090 graphics card this time, the chip behind which has the extremely catchy name, GB202.

We're not going to go through everything with a fine-tooth comb, because that would make this review 10 pages long. Nvidia has chosen to focus more on AI, as we are already down to such small sizes when it comes to hardware that it is problematic to push things further. At these levels, Diminishing Return is a serious problem. Ideally, I'd actually like to go through the whole thing because it's technologically extremely impressive, but for anyone other than the ultra-hardcore tech geeks, it's also just nonsense talk, but impressive nonetheless.

Let's start with the design. Nvidia has kept its heavily industrial and feature-based design, which I love, and has now moved to a dual-blow-through design. That is, a PCB with the chip in the centre and symmetrically arranged cooling pipes on either side, each with its own fan that draws air through the card. It seems like a more logical way to do it, and it's a little hard to understand the many intermediate steps between this solution and the old design.

It also means that the card is now a traditional 2-slot design, which has helped the size enormously. Unfortunately, it's still over 30 centimetres long, which is still too much in my opinion. Nvidia has even claimed that this Founder's Edition can fit into small SFF enclosures that typically manage 33-35 centimetres, but I wouldn't try that myself.

In addition, the power connector does not extend perpendicularly from the card, but is angled and thus protrudes significantly less than before. A new power adapter has been designed with much more flexible cables that are not braided or extremely stiff, which also helps. Speaking of adapters, I do recommend a dedicated cord if you have one. The attentive reader will realise that there are now not three, but four connections for 8-pin power cables. The good news is that if you use the right connectors, you can actually make do with three, but you do lose some performance, but typically only 10-15%. The bad news is that you need all four to get full value for money, resulting in 575.8 Watts being drawn through the system. Yikes. It's a good thing modern CPUs have become more energy efficient, because that's a lot of power. The GPU runs slightly above the promised 2407 MHz, and the same goes for the RAM speed, which officially stands at 2209 MHz.

But back to the internals. We're now up to 32 GB of GDDR7 VRAM in the beast. That's insane. A 4nm TSMC production platform, over 92 billion transistors, and PCIe 5.0 connectivity actually comes into play. We hit a VRAM bandwidth of 1.79 TB/s. On top of that, you get 21,769 Shading Units (aka Stream Processors - they're the ones doing most of the work, roughly speaking) - which now support Neural Shading and thus generative AI down to this level. Theoretically, this allows real-time graphics of, among other things, faces with a much higher level of light, shadows, and reflections than before, and the realism of the graphics should increase significantly if game developers can allocate resources to utilise it. There are a lot of other innards, but for most people the most relevant will be that there are 170 Ray-Tracing cores and 680 Tensor cores of the latest generation, with these responsible for everything related to upscaling and Deep Learning.

The outputs include one HDMI 2.1b and three DisplayPort 2.1b. It's a bit strange that the HDMI port isn't 2.2 when you're launching something with a DisplayPort standard that's so new that I can't immediately find anything else that has it, perhaps because the standard isn't technically finished yet. There's also support for 4K/480 Hz or 8K/165 Hz with some compression. It'll probably bite me in the arse, but 4K at 240 Hz should be plenty for even the most demanding gamer, although there will probably be esports players who want more of everything.

If we had to highlight one thing in particular, apart from the fact that those who use GeForce graphics cards for professional use will probably want to buy one because they've upgraded not only the processing power but also the encoding and decoding dramatically, that would be DLSS 4. They've switched to a completely new way of upscaling and have switched to the so-called Transformer model, which in practice means much less ghosting and far fewer artefacts. In fact, I would go as far as to say that DLSS 4 is a big step forward. Then there's the more controversial part, namely Multi Frame Generation. Where we used to get one artificial image between two "real" ones, we now get up to three. This has led many to criticise the "fake" images, but what is perhaps forgotten is that the latest generation of graphics cards allow much more to be done in the graphics card and not in the processor itself in the computer.

More of the graphics in modern games will also be generated directly in the game instead of being an existing data set to be calculated on. This requires a lot of computing power, but the delays are minimised with the Reflex 2 system. However, we have to get used to a completely different way of thinking about graphics generation, especially with the many light and shadow requirements, and the fact that the graphics card thinks more on its own. We will still show benchmarks with native results, without any voodoo switched on.

Before we get to that, we need to get one thing straight: the price. A top-of-the-range model like the one we've borrowed here doesn't come cheap. It's £1,939/$1,999, and the cards you can buy in stores that aren't made by Nvidia will probably cost more. It's a scary and insane price that also clearly indicates that this card is not intended for mainstream consumers, which is ironic as it's aimed at those who game in 4K on fast monitors with 200+ Hz refresh rates. Things get a little more fun with the RTX 5080 card, which "only" costs £979/$999, and the two even cheaper RTX 5070 cards (around £539/$549). They may not have quite the same 4K capability, but you'll get all the goodness of the RTX 50 Series at a price where many more can suddenly join in, relatively speaking. But a price increase of a few hundred pounds from the RTX 4090 in less than a year and a half can't be excused by inflation or anything else, it's just pure madness. But buyers for Nvidia's XX90 cards have never been like the rest of us either.

And now for the fun part: benchmarks. We tried to get them right as much as possible with the drivers that were available, on an X870E platform, and with everything running on a PCIe 5.0 platform for both drivers and graphics cards.

The heat output was quite surprising: it idles at 48 degrees Celsius, maxes out at 72 degrees Celsius, but most of the time it was holding at 63 degrees Celsius. It should also be said that the test bench is designed for maximum airflow, but it's impressive regardless. The noise hit 41.5 dB, but is extremely low frequency. No coil whine, no fan blade noise or, worst of all, air turbulence. So it's a job well done.

The performance - which with DLSS 4 is double that of the RTX 4090 - is still, in raw processing power, quite impressive. It's double the performance of the RTX 4080 Super and in some tests even more. There are also new and even wilder Ray-Tracing modes for several games, but as we were unable to collect data on these with older cards, they have not been included.

3D Benchmark

• Time Spy: 36332


• Time Spy Extreme: 18503


• Speed Way: 14251


• Port Royal: 36188


• Steel Nomad: 14050

Here we see a massive improvement in most areas, with Port Royal being a 50% improvement from the RTX 4090.

V Ray 6

• RTX: 15286

V-Ray 6 is a rendering programme that has a solid benchmark tool and a dedicated mode for several types of rendering for both CPU and GPU. The score is three times what an RTX 4080 Super can do, but far from the world record set by the 15x RTX 4090. Yes, it's for professional use. Black Myth: Wukong has a nice benchmark tool that really shows its worth: 4K, everything on maximum, it gives 32 FPS. 56 with Frame Generation. We'll be using this benchmark for a long time. And then for gaming!

Total War: Warhammer III

• 1080p: 346.00


• 1440p: 270.20


• 4K: 161.30

Total War still doesn't use any assistive technology and is tough on the CPU and GPU, but still offers a 60% improvement over the RTX 4090.

Red Dead Redemption 2

• 1080p: 207.22


• 1440p: 189.85


• 4K: 133.42

Cyberpunk 2077

Ultra settings, Ray-Tracing Ultra/ DLSS 4 + Multi Frame Generation:

• 1080p: 154.94/548.78


• 1440p: 157.48/488.99


• 4K: 128.36/347.05

Assassin's Creed Valhalla

• 1080p: 235.00


• 1440p: 199.00


• 4K: 145.00

Far Cry 6

• 1080p: 192.00


• 1440p: 183.00


• 4K: 156.00

Both Assassin's Creed and Far Cry 6 are unfortunately showing their age, it still takes a lot to pull it off in 4K, but as for the rest, it doesn't really matter.

Metro Exodus - Enhanced Edition

4K Extreme, without and with DLSS:

• 4K: 96.50


• 4K with DLSS: 184.16

Despite DLSS 3 making a significant difference here, it's the fact that it can run at almost 100 FPS that you should take note of. This is a test that unfortunately often crashes many graphics cards, and even expensive models rarely pull over 50 FPS.

Star Wars Outlaws

4K Ultra Quality, without and with DLSS:

• 4K: 104.00


• 4K + DLSS: 311.00

Here was a really successful implementation of DLSS 4. No artefacts, strange glitches, ghosting, or anything else, it was virtually impossible to tell the difference between native and Frame Generation.

In the meantime, we have:

Dragon Age: The Veilguard

4K Ultra Quality, full Ray-Tracing:

• 4K: 84.00


• 4K + DLSS + Multi Frame Generation: 181.00

Here was also an extremely successful integration where it was virtually impossible to tell the difference between native and Multi Frame Generation.

Call of Duty: Black Ops 6

Without and with DLSS switched on:

• 1080p: 245.00/431.00


• 1440p: 205.00/352.00


• 4K: 146.00/245.00

Warhammer 40,000: Space Marine II

4K, without and with DLSS switched on:

• 4K: 107.00/160.00

So, here comes the big question: should you buy one? That's a bit complicated. DLSS 4 is a bit of a game changer. If you're playing around with quality settings, you may already be familiar with Ray Reconstruction and DLAA, but the transition to the Transformer model when upscaling, i.e. using DLSS, is very visible. If you're reasonably happy with your RTX 40 Series card, it may not make much sense to switch, especially as rumour has it that there will be up to a 10% performance improvement just by switching to this (again a rumour) for RTX 40 Series cards.

Conversely, Multi Frame Generation is only available in the RTX 50 Series, and it seems to work impressively well, especially when combined with a card powerful enough to pull everything in 4K native, which it should for the price. Unfortunately, the price is probably my biggest problem. If they had raised the price a bit or kept it around the same level, I would have been happy, but an increase of almost 20% is too much. Even if it's much more powerful than its predecessor because you expect that with the absolute flagship model.

On the other hand, if you have an older card, such as the RTX 3090, then yes, you should probably upgrade if you want to play in 4K. More and more of the games we look at here at Gamereactor are poorly optimised and incredibly poorly adapted to the PC, so a lot of horsepower must be available all the time. However, Nvidia should be commended for listening to the criticism and making a better power connection and a smaller card. But four 8-pin connections? It's still completely unreal to look at.