NVIDIA GeForce RTX 5090

Written by John Yan on 1/25/2025 for PC  

Announced at CES 2025, NVIDIA ushered forth the Blackwell era with four announced video cards. The main halo product, of course, is the XX90 SKU and today, we’re going to look at the flagship NVIDIA GeForce RTX 5090.

As you can see, the GeForce RTX 5090 is a dual slot card, which is truly a marvel of engineering considering this card has a higher TGP than the three slot GeForce RTX 4090. Not only is the card taking up less slots, the PCB is smaller than the RTX 4090 encompassing the center of the card between the two blow through fans. But let’s step back for a sec.

Blackwell is the name of the architecture for this generation’s NVIDIA line. Announced at this year’s CES, AI has a big influence on the Blackwell architecture, and it’s not hard to see why. When used appropriately, it can be a great time saver. Because of that, there’s an AI Management Processor included with the Tensor Cores, RT cores, and SM cores.

Neural networks have been integrated into the programmable shaders producing what NVIDIA has dubbed as neural shaders. What can be done with these neural shaders? How about improving textures and advanced lighting effects? One thing that might be hugely beneficial is being able to compress textures by up to 7X. When we’re talking smaller sized games but still with great visuals, texture compression can help.

What NVIDIA really highlighted at CES and their keynote was DLSS4. Their new transformer model for Ray Reconstruction and Super Resolution is more powerful and provides better performance and image quality over the previous models.

The big thing though seems to be Multi Frame Generation. DLSS 3 featured Frame Generation where the video card would generate every other frame. DLSS 4 now leverages AI to generate up to three additional frames. So if you use DLSS 4 at its highest level, you’d see one game generated frame and three additional video card generated frames. You can set it to generate every other frame, two additional frames, or three additional frames. So if you want the old way of Frame Generation, that’s possible. I think this is one of the more divisive aspects of the Blackwell architecture and this might be another feature that will change people’s mind as time goes on. I know when NVIDIA first introduced Frame Generation, there was a big divide on if this was a good option or not. Multi Frame Generation, I think, will probably also go through that same period until people are more exposed to it. I think we’ll know a lot more in a year or so to see if Multi Frame Generation will be something that more people accept and use.

So now we got the flagship GeForce RTX 5090, which will retail for $1999. Spec wise, the RTX 5090 has 21760 CUDA cores. That’s an increase of 32.8% in count and, of course, a generational leap. More Tensor Cores were put in as well with an increase of 168 Tensor Cores totalling 680 on the RTX 5090. Base and boost clock does go down a tad with the base clock set at 2010 MHz and the boost set at 2410 MHz.

VRAM has been increased over the RTX 4090 as the RTX 5090 has 32GB of GDDR7 memory. That’s 8GB more than the RTX 4090. I really do wish NVIDIA would increase the amount of VRAM across the board with the Blackwell line though, but it is what it is. With the extra 8GB of VRAM, the GeForce RTX 5090 will be capable of loading in larger AI models so if you’re like me and run some local LLMs, this is a nice improvement over the previous XX90 cards. The memory bandwidth is about 78% more than the RTX 4090 at 1792 GB/s.

TGP or total graphics power has gotten a significant bump as the TGP for the GeForce RTX 5090 is 575W. Compared to the RTX 4090, which had a TGP of 450W, it’s a 27.8% increase in possible total power draw. NVIDIA recommends that you have a 1000W or greater PSU, so if you plan on getting this card and don’t have that, be sure to add that to the upgrade list when picking up this card.

Spec wise, here’s how it stacks up to both the RTX 4090 and RTX 3090 Ti:

	GeForce RTX 5090	GeForce RTX 4090	GeForce RTX 3090 Ti
SM	170	128	84
CUDA Cores	21760	16384	10752
Tensor Cores	680 (5th gen)	512 (4th gen)	336 (3rd gen)
RT Cores	170 (4th gen)	128 (3rd gen)	84 (2nd gen)
Texture Units	680	512	336
ROPs	192	176	112
Base Clock	2010 MHz	2235 MHz	1560 MHz
Boost Clock	2410 MHz	2520 MHz	1860 MHz

As mentioned before, the GeForce RTX 5090 is a dual slot card. While it’s still the same length as the RTX 4090, it does take up one less slot than the previous generation halo product. That should make some SFF folks happy as they can slot this into smaller cases.

Inside, the card is now split up into four separate pieces. That’s a big change from the entire PCB holding everything together and having a cooler placed on top of it. For the Founder’s Edition, the RTX 5090 features a main PCB board with all the memory, GPU, and hardware to drive your graphics. The I/O that houses the four video connectors is its own separate daughter board as well as the PCI-E slot. Finally, a flexible cable connects the I/O daughter board to the main PCB. The PCI-E daughterboard connects to the main board directly. Making it modular must have taken some great engineering to accomplish this and I wonder if this is how we’ll see video cards done moving forward.

The looks of the card is similar yet different from the previous generation. There’s still the two fans, but now both are on the front and blow through the card instead of being opposite spinning fans in the previous few generations. There’s a few great videos on other content creators on NVIDIA’s decision to go this route as well as some of the other improvements such as using liquid metal on the core to better transfer heat to the cooler as well as why there’s a concave design on the fins, which now runs vertically on the card rather than at a 45 degree angle and horizontally.

NVIDIA is still using the 12VHPWR connector, but now the placement is a lot better. Rather than being straight up and parallel to the back of the card, the connector now sits at an angle and perpendicular to the back. This does allow for an easier bend around the top of the card and to the back of the case. Hopefully, this lessens the stress on that port and with the increase on TGP, anything to help make the power connector less problematic the better. If you don’t have a 12VHPWR cable coming from your PSU, NVIDIA includes a 4-1 power adapter which is designed to be a lot more flexible. Rather than cable sleeves, it’s now a more stranded design which makes it a lot less stiff.

There’s still the familiar white LED coming from the inner part of the card where the tips of the triangles are and the glowing Geforce RTX logo on top. Colorwise, I do like the more muted, darker grey color and darker accent piece.

I/O has changed a little bit here and for those complaining about the lack of DisplayPort 2.1 ports, that’s been fixed. We will have three DisplayPort connectors and one HDMI connector on the Founder’s Edition card, but each port has been upgraded to the latest available standard. There were people who didn’t like that the previous generation card didn’t come with DisplayPort 2.1, but here NVIDIA went and put in DisplayPort 2.1b, which is a throughput of 80 Gbps.

HDMI has gone from HDMI 2.1a to HDMI 2.1b. It’s not much of a change from 2.1a though, but does tidy up some of the standards for the port. HDMI 2.2 is the next major change, but products that support that won’t be out for a few more months.

Earlier this week, I posted pictures of the unboxing and the new eco-friendly packaging that the GeForce RTX 5090 comes in. It’s a good design in my opinion and the entire packaging will be easy to recycle.

So let’s see how this card performs against the previous generations’ top dog, the GeForce RTX 4090. Here’s what I have for my test system:

AMD Ryzen 7 9800X3D
G.Skill Trident Z5 Neo RGB 64GB (2 x 32GB) DDR5-6000 PC5-48000 CL30
ASRock B65-E PG Riptide WiFi motherboard
NVIDIA GeForce RTX 5090
Samsung 990 Pro 4TB SSD
LG OLED42C2PUA 42"
NVIDIA Drivers 571.86

Let's start with some basic rasterization figures. All tests in this review were done at 4K resolution.

Depending on the game, we’re seeing anywhere from it being slightly slower in a CPU bound game like Microsoft Flight Simulator to a 53% increase in a game like Cyberpunk 2077. Those two are the extremes though as we seem to get an average of about 30% or so once you take all the games into account. It’s not that huge uplift that we saw coming from an RTX 3090 to an RTX 4090, but there is a performance increase nevertheless in rasterization.

Let’s talk about DLSS. NVIDIA introduced DLSS4 with Multi Frame Generation and a new Transformer Model for upscaling and ray reconstruction at CES 2025. The new Transformer Model is supposed to bring forth much better image quality when using DLSS. Below is a video that helps explain how it works and compares it to the previous way of doing DLSS along with Multi Frame Generation.

Having witnessed it at CES in person, the new Transformer Model really does improve DLSS quality. Things such as ghosting have either been eliminated or minimized in some of the scenes shown and there were areas with higher quality looking textures that made the scene much better looking. What’s nice is you don’t need a Blackwell card to take advantage of this and you can use this feature right away all the way as far back as the 20 series.

 

I tried running DLSS on Microsoft Flight Simulator 2024, but the game didn’t get any performance improvement in any setting in 4K. That said, you can see how much better the RTX 5090 runs games using DLSS and Cyberpunk 2077’s performance really shows using the new Transformer Model. The image quality was fantastic and a lot of the issues such as ghosting and shimmer improved dramatically over the previous method for doing upscaling. There’s still some issues though like in Star Wars Outlaws where a grate was shimmering when looked at from a certain angle with DLSS on. I’m guessing that’s using the older method as there wasn’t a way to choose which DLSS method to use like in Cyberpunk 2077. If the other games perform like how Cyberpunk 2077 did with the new Transformer Model, I’m excited to see other games utilize it.

Now let’s turn on ray tracing where applicable. Some games feature full path ray tracing like Indiana Jones and the Great Circle as well as Cyberpunk 2077.

Again, some nice uplift in performance over the RTX 4090 without getting into the Multi Frame Generation or upscaling DLSS4 provides. But let’s see what happens when we throw in Frame Generation in at 2X along with DLSS features enabled with RTX.

As we’ve seen in the past, Frame Generation can really smooth out the performance of a game. It does come at a small cost of latency though and I think for slower style games like simulators and single player games, that’s fine. I wouldn’t turn it on for anything competitive. But, let’s dive into the newest feature of DLSS4 and that’s Multi Frame Generation.

I talked a little bit earlier about Multi Frame Generation so I won’t go into it too much again. Games can use Multi Frame Generation in two ways. There’s, of course, built-in support like what we see in Cyberpunk 2077’s recently released patch. The other is using the NVIDIA app to force it to be in use. Options will be available to enable it this way should your game not have built in support.

So, how well does it do? Let’s see the numbers in Cyberpunk 2077 and Star Wars Outlaws.

Who'd thought we'd see Cyberpunk 2077 with everything turned on running at close to 250 frames per second? Now, I wouldn't call this a better performance result, but more like a better smoothing result. Would I turn it on for Cyberpunk 2077? As a single player game, yes. There’s artifacting, but only if you really look for it. I think with how the graphics become super smooth with path tracing at 4K, it’s a feature I’d have no issues using. Star Wars Outlaws was the other game I tested with this feature on and it worked pretty well as well. Now, I’ve seen some reviewers say they experienced flickering or popping textures, but I think that’s more so on the game than the NVIDIA tech. I’ve seen some posts in the past saying you need to edit some lines in a config file to fix this and it’s a bug of the game. I was able to get some missing floor textures to happen in some areas during some testing and it was fine at other times and this happened on straight rasterization tests. That said, the increased smoothness in Star Wars Outlaws was pretty awesome to see with 4X Frame Generation.

Finally, we have some performance in VR and 3D Mark Time Spy Extreme at 4K.

Multi Frame Generation is something that, I think, we will need time to assess if this technology really is a good way to go. Having the old way of generating every other frame took a bit to be accepted. Now you’re asking the public to be OK with the video card outputting up to two more additional generated frames for a total of three. The good news is that you can control how many additional frames you want the card to generate so you aren’t going to be forced to 3X or 4X Frame Generation if you don’t want to. But, NVIDIA is really highlighting this and it’s one of the main selling points of the Blackwell architecture as this feature isn’t available on previous generation cards, unlike the new Transformer Model.

Is this card a monster? Yes. Is it worth an upgrade over an RTX 4090. I would say, no unless you really want to use DLSS4 and/or work in AI as the higher VRAM capacity will benefit those that load in AI models locally. The engineering to get it down to a 2 slot card is pretty impressive given the increase in TGP. It definitely does use more power as my UPS did increase about 100W or so over the RTX 4090. We aren’t seeing that huge performance uplift in traditional rasterization and RTX scenarios like we did from the RTX 3090 to the RTX 4090 unless you factor in DLSS4 Multi Frame Generation. As some folks I overheard when I was at CES’s NVIDIA booth, this is gonna be a weird generation to test and it's one of the ones where I’m not confident yet to know which way I fall on these new features. Traditionally it’s all about rasterization performance versus the prior generation. But with the advent of new DLSS AI models and Multi Frame Generation support among other things, we have a lot more data points to look at when comparing video cards.

The 25% increase in price over the RTX 4090 does translate, mostly, to a straight 25-35% average increase in performance in certain situations not factoring in DLSS4. So cost wise, it’s not getting more bang for the money you spend. But then again, the XX90 cards were always going to be priced at a premium. I think it’ll be more interesting to see how the other cards announced perform over its predecessor and we’ll have that as the releases happen.

But for now, the GeForce RTX 5090 is a monster performer of a card when looked at on its own. Supplies will be limited however and you’ll probably be hard pressed to find it at the $2000 price point.