GeForce RTX 5090 Founder’s Edition Review – Gamer’s Pricey Dream

In just a few hours, NVIDIA will officially launch its next flagship graphics card, the GeForce RTX 5090 Founder's Edition. There might well be shortages or price hikes, but that's another matter entirely. Hardcore PC gamers have been waiting for more than usual (the GeForce RTX 4090 launched in October 2022) to get a new GPU, and they're eager for it despite the high price (the graphics card's official MSRP is $1,999), as shown by the queues that have already started forming outside some stores. Should you be paying that much, though, even if you can afford it? In this article, I'll tackle that question after testing the graphics card strictly from a gamer's point of view.

First of all, some pictures of the hardware component and the package it came in. The GeForce RTX 5090's design is certainly very sleek, and I appreciate that it takes up two slots rather than three.

If you're a regular Wccftech reader, you're probably well-informed on all the specs and features of the GeForce RTX 5090 through our hardware coverage. Still, a quick recap is in order. This is a monster GPU with 21760 CUDA Cores, 32 GB GDDR7 VRAM, 512-bit memory interface, 3352 AI TOPS for Tensor Cores operations, and 318 TFLOPS for Ray Tracing operations. With all that, it's no wonder it requires 1000W of system power, drawing up to 575W for itself.

On the productivity side, the RTX 50 Series includes hardware support for encoding and decoding the 4:2:2 pro-grade color format, as well as 9th Gen NVENC video encoders and a 6th Gen NVIDIA decoder. NVIDIA also added support for the latest DisplayPort 2.1b UHBR20 specification, which can go up to 8K@165Hz with Display Stream Compression (DSC).

GPU Engine Specs	NVIDIA CUDA Cores	21760
	Shader Cores	Blackwell
	Tensor Cores (AI)	5th Generation 3352 AI TOPS
	Ray Tracing Cores	4th Generation 318 TFLOPS
	Boost Clock (GHz)	2.41
	Base Clock (GHz)	2.01
Memory Specs	Standard Memory Config	32 GB GDDR7
	Memory Interface Width	512-bit
Technology Support	NVIDIA Architecture	Blackwell
	Ray Tracing	Yes
	NVIDIA DLSS	DLSS 4 Super Resolution DLAA Ray Reconstruction Frame Generation Multi Frame Generation
	NVIDIA Reflex	Reflex 2 Low Latency Mode Frame Warp (Coming Soon)
	NVIDIA Broadcast	Yes
	PCI Express Gen 5	Yes
	Resizable BAR	Yes
	NVIDIA App	Yes
	NVIDIA Ansel	Yes
	NVIDIA FreeStyle	Yes
	NVIDIA ShadowPlay	Yes
	NVIDIA Highlights	Yes
	NVIDIA G-SYNC	Yes
	Game Ready Drivers	Yes
	NVIDIA Studio Drivers	Yes
	NVIDIA Omniverse	Yes
	RTX Remix	Yes
	Microsoft DirectX 12 Ultimate	Yes
	NVIDIA GPU Boost	Yes
	NVIDIA NVLink (SLI-Ready)	No
	Vulkan 1.4, OpenGL 4.6	Yes
	NVIDIA Encoder (NVENC)	3x Ninth Generation
	NVIDIA Decoder (NVDEC)	2x Sixth Generation
	AV1 Encode	Yes
	AV1 Decode	Yes
	CUDA Capability	12.8
	VR Ready	Yes
Display Support	Maximum Digital Resolution	4K at 480Hz or 8K at 165Hz with DSC
	Standard Display Connectors	3x DisplayPort, 1x HDMI
	Multi Monitor	up to 4
	HDCP	2.3
Card Dimensions	Length	304 mm
	Width	137 mm
	Slot	2-Slot
	SFF-Ready Enthusiast GeForce Card	Founders Edition - Yes Varies by manufacturer
Thermal and Power Specs	Maximum GPU Temperature (in C)	90
	Total Graphics Power (W)	575
	Required System Power (W) (5)	1000
	Supplementary Power Connectors	4x PCIe 8-pin cables (adapter in the box) OR 1x 600 W PCIe Gen 5 cable

The GPU is the pinnacle of the brand-new Blackwell architecture, which NVIDIA says is optimized for neural rendering. Doubling down on AI, NVIDIA has created the so-called RTX Neural Shaders (part of the larger RTX Kit suite), small neural networks that can be embedded into programmable shaders. An RTX Neural Shaders SDK lets developers train their game data and shader code on an RTX AI PC and accelerate their neural representations and model weights at run-time with NVIDIA Tensor Cores.

So far, NVIDIA has identified three key applications: RTX Neural Texture Compression, RTX Neural Materials, and RTX Neural Radiance Cache. However, all those will require support from the graphics APIs; Microsoft has already said that the Cooperative Vectors are coming soon to DirectX 12. The Khronos Group will undoubtedly come up with a similar solution for Vulkan. Afterward, though, it will be up to game developers to implement these features in their games. Long story short, it might be a while before we see RTX Neural Shaders applied to actual games, making it automatically not a focus of the GeForce RTX 5090 launch reviews.

The same goes for NVIDIA Reflex 2; the new Frame Warp technology promises to cut down PC latency by up to 75% (the original Reflex reduced latency by up to 50%), which sounds great, especially for Multi Frame Generation since it slightly increases latency. But Reflex 2 isn't available yet - and it won't be exclusive to RTX 50 graphics cards anyway.

What is necessarily at the front and center of every conversation surrounding the new RTX 50 GPUs is, of course, NVIDIA DLSS 4. Now, there's two parts to that: on one hand, NVIDIA has released a new, more advanced upscaling model based on vision transformer technology. So far, DLSS has been based on traditional CNN (convolutional neural network) models. According to NVIDIA, though, these models simply aren't smart enough to make the best choices, which sometimes translates into visual artifacts or ghosting. The new DLSS 4 transformer models are trained on twice the amount of parameters and, according to NVIDIA, already provide better results than CNN models, not to mention they can be further improved over the coming years, whereas CNNs have pretty much reached their maximum capabilities.

The new Ray Reconstruction model already improves image quality, while the Super Resolution model is being released as a beta.

The new models are available to every GeForce RTX user. Similarly, the new Frame Generation model, which replaces the hardware Optical Flow accelerator with a faster and more efficient AI-based model, is available on both RTX 40 and 50 GPUs.

What's exclusive to Blackwell architecture graphic cards like the GeForce RTX 5090 is DLSS Multi Frame Generation. According to NVIDIA, there's a number of hardware advancements making generating multiple frames possible. The main one is the 5th Gen Tensor Cores, which reportedly offer up to 2.5X more AI processing performance. Moreover, the new graphics cards have been enhanced with hardware Flip Metering capabilities for optimal frame pacing. Previously, Frame Generation relied on traditional CPU-based pacing, which can introduce variability that affects smoothness. With the arrival of DLSS 4, NVIDIA has updated Frame Generation and Multi Frame Generation to use Flip Metering instead of CPU-based pacing, allowing the display engine for 40 and 50 Series GPUs to precisely control display timing.

With DLSS Multi Frame Generation, NVIDIA is promising absolutely massive performance gains of up to 8x in the heaviest games, like Alan Wake 2, Black Myth: Wukong, and Cyberpunk 2077.

However, those best-case scenario frame rate uplifts are measured with the DLSS Performance Mode; in this review, I've instead put the GeForce RTX 5090 through the grinder when using the highest quality DLSS setting available in any game, DLAA or, if that's not an option, Quality Mode.  In all cases, I've enabled the maximum in-game settings without resorting to out-of-game tweaks or mods. 4K resolution has been selected in all instances.

The testing rig is:

• AMD Ryzen 7 9800X3D
• MSI MAG B650 TOMAHAWK WIFI
• 32GB DDR5 6000 MHz G.SKILL Trident Z5
• 2TB NVMe SSD Western Digital Black SN850X

Before starting, it's worth clarifying that all testing was conducted with the latest (not public yet) version of NVIDIA's FrameView software. There's a very good reason for that, as it is presently the only benchmarking software that accurately reports the MsBetweenDisplayChange metrics. With the aforementioned new Flip Metering, NVIDIA says that MsBetweenDisplayChange, rather than MsBetweenPresents, is the best metric to represent what the user actually sees. That's because it is reported near the end of the render pipeline after flip metering has occurred, whereas the measurement of MsBetweenPresents takes place earlier when flip metering hasn't yet paced the frames.

As advised by NVIDIA, I exported FrameView data to be visualized with FLAT (Frame Latency Analytics Tool), which was then used to create the frame time pie charts.

The first game to be checked out couldn't be anything other than Cyberpunk 2077. CD Projekt RED has constantly paved the way forward for NVIDIA technology, first by delivering several ray tracing options at launch, then by adding DLSS Frame Generation (3.0), Ray Reconstruction (3.5), path tracing, and so on. Now, it's time to see how it fares with the new NVIDIA DLSS 4, which the game was the first to implement thanks to Update 2.21.

 

The transformer model costs little in Cyberpunk 2077. Its performance is 94% that of the previous CNN model, so when you have frames to spare, it is well worth keeping. And you do, thanks to the incredible power of the GeForce RTX 5090. With literally everything set to max, path tracing on, and DLSS set to DLAA (native 4K resolution), we're at 120 frames per second.

Moving onto frame times, FLAT shows 6.10% of stuttering time, 6.08% of 'variable FPS', and 87.82% of consistent FPS. More importantly, the game felt smooth as butter.

When it comes to latency, the new transformer model registers a negligible penalty compared to the CNN model. Overall, the latency is roughly the same when you use Frame Generation in Cyberpunk 2077.

Like CD Projekt RED, Remedy was an early proponent of ray tracing, dating back to the release of Control. With Alan Wake 2, they've stepped it up, providing full-blown path tracing with their proprietary Northlight engine.

On the new GeForce RTX 5090 Founder's Edition, the game looks and runs absolutely great. It should be noted that the tested build features multiple improvements over the public build. There's a new Ultra ray tracing preset (which is not currently selectable when using AMD or Intel GPUs), which adds fully ray traced refractions, fully ray traced transparent reflections, and enhances the quality of the fully ray traced indirect lighting. Moreover, this build already includes RTX Mega Geometry, which NVIDIA says helps reduce CPU, GPU, and VRAM usage when ray tracing is enabled and also improves its quality because all raytracing details are now updated with every frame, reducing artifacts.

Even when enabling the new DLSS transformer model in 4K Quality Mode, the average frame rate stayed well above 150 FPS, keeping 92% of the performance registered with the CNN model. However, I do have to say that even the new DLSS model does not completely remove some ghosting trails that can be seen when quickly moving the camera, such as in the embedded video.

There is a tiny bit more stuttering than in Cyberpunk 2077 (8% instead of 6%), but the in-game experience is still very smooth, thanks to the updated Flip Metering.

Once again, the latency does increase a little when switching NVIDIA DLSS 4 to the transformer model, but not to a degree that could be actually felt when playing.

Out of all the games that NVIDIA made available to press members when reviewing the GeForce RTX 5090 Founder's Edition, Marvel Rivals is by far the least taxing. Sure, it is powered by Unreal Engine 5, but it is an arena-based shooter designed for fluidity.

With the new beastly GPU made by NVIDIA, users can reach ridiculous average FPS results even with the game maxed out (including Lumen GI) and DLAA enabled. We're talking over 300 frames, though admittedly, I was strolling in the Practice Range map during the test to achieve consistency. It might be a little lower during more chaotic multiplayer matches, but the experience is nonetheless top-notch and can fully take advantage of the latest high refresh rate displays.

Here, the transformer model stays at 96% of the average FPS registered by the CNN model.

Only 4% of the frame time is reported under stuttering, 4% under FPS variability, and nearly 92% falls under the average FPS label.

With such high frame rates, it's no wonder to see low latency. Here, CNN and transformer models are essentially tied.

BioWare's Dragon Age: The Veilguard, powered by the Frostbite engine, ran fairly well at launch with NVIDIA DLSS Super Resolution and Frame Generation. However, the ray tracing effects were a bit heavy even on an RTX 4090, and Super Resolution had to be dropped down a few notches to make things smoother.

The GeForce RTX 5090 Founder's Edition levels up the whole experience. With DLAA enabled, ray tracing enabled, and the new DLSS 4 Multi Frame Generation set to x4, the game broke 250 average frames per second when roaming through Tevinter's Old Town map in an impressive performance display.

Unfortunately, since Dragon Age: The Veilguard was tested with the NVIDIA app Override functionality (the game itself hasn't been updated yet with the new features), I couldn't compare the old CNN model with the new transformer one.

The improved frame pacing also diminishes the stuttering instances. The data shows only 6.56% stuttering time, 3.42% variable FPS, and 90% stable FPS.

NVIDIA DLSS 4 Multi Frame Generation does increase latency a little when set to x4 compared to x3 or x2, but the amount is practically imperceptible, whereas the average frame rate benefits of using x4 are vast.

Star Wars Outlaws by Ubisoft Massive is one of the most visually impressive games you could play at this point in time. It's an excellent showcase of the Snowdrop engine and delivers a fantastic rendition of Star Wars.

While not path traced, the game offers multiple ray tracing options, the most expensive of which is RTX Direct Illumination. When I played the game with my RTX 4090 back at launch, I had to disable RTXDI and turn DLSS down to Performance mode. The RTX 5090, however, allowed me to activate DLAA, re-enable RTXDI, and just enjoy Star Wars Outlaws in its peak visual splendor at over a hundred average frames per second.

Out of all the games, Star Wars Outlaws performed the best when measuring frame times. Stuttering was practically non-existent at just 2.43%, with 93% stable FPS and 4.5% variable FPS. It really felt incredible when playing it without all the hitches that I first experienced just a few months ago.

On the flip side, Star Wars Outlaws recorded the worst latency of all at over 80 milliseconds. The controls still felt quite responsive, though, enough for a single player game that doesn't rely on reflex that much.

The best-selling game of 2023, Hogwarts Legacy, is getting a fairly significant tech update. Not only is it adding support for DLSS Multi Frame Generation and the new DLSS 4 transformer models, but the game's iffy default ray tracing (which various mods tried to improve, as you might recall) is being enhanced by the addition of Ray Reconstruction. Moreover, all ray traced reflections are now rendered at a higher fidelity, resulting in more detailed visuals, and the reflections have also been enabled on water bodies, including the Black Lake.

Needless to say, the game has never looked this good. It also runs at very high frame rates, at least when looking at the average FPS. However, the difference between average FPS and 1% Low FPS is huge, particularly for the x4 MFG mode.

In practice, this means there's more stuttering in this game than in any other we've tested for this review. This is confirmed by the FLAT pie chart below - 13% of the frame time is reported under stuttering.

Now, it should be noted that Hogwarts Legacy, just like many Unreal Engine 4 games released at that time, has always been prone to heavy stutters, particularly when roaming within the school itself. This is valid to this day with any configuration, and not even the combination of GeForce RTX 5090 and DLSS Multi Frame Generation can fix that issue. It needs to be resolved from the developer's side. I certainly hope they'll do it for the rumored Definitive Edition.

This wasn't the only issue I encountered with Hogwarts Legacy, which also exhibited noticeable ghosting around the main character's head. The transformer model is an improvement, to be sure, but there's definitely room for further refinements.

When it comes to measured latency, Hogwarts Legacy's latency is around 70ms, which is a similar value to Cyberpunk 2077 and Alan Wake 2.

Those were the games that were updated to support the new NVIDIA DLSS 4 Multi Frame Generation ahead of the official launch. However, I still tested the GeForce RTX 5090 with a few of the most taxing PC games available, starting with Game Science's Black Myth: Wukong.

The stunning Unreal Engine 5 powered action RPG features full-fledged path tracing and is really heavy on any PC. With the benchmark tool, handily available via Steam as a separate app, the new flagship GPU shows its prowess, reaching nearly a hundred frames per second with just the regular Frame Generation. Once the game is updated to support Multi Frame Generation, I can easily see it running at over 200 frames per second on this beast of a GPU.

While a bit older, Dying Light 2 Stay Human features several ray tracing options (shadows, ambient occlusion, reflections, global illumination, and flashlight). With DLSS set to Quality and everything maxed out, the GeForce RTX 5090 crushed the benchmark with over 210 average frames per second.

Another good old benchmark is IO Interactive's HITMAN World of Assassination (formerly known as HITMAN 3), powered by the latest version of the Glacier engine and enhanced with ray traced sun shadows and reflections. I picked Dartmoor, the physics-heavy benchmark level, which again reported very, very high average FPS, although the 1% Low FPS was a bit lower than I expected.

The acclaimed Unreal Engine 4-based roguelike third-person shooter Returnal features a really detailed benchmark, which, again, reported well over 200 average frames per second when run on the GeForce RTX 5090 Founder's Edition, even with ray traced moon shadows and reflections.

Arguably the first heavily ray-traced game, Metro Exodus Enhanced Edition is still a noteworthy benchmarking test. The new NVIDIA flagship graphics card surpassed 130 average frames per second, and even the 1% Low FPS was nearly 100 frames per second—all this without needing Frame Generation, which isn't supported.

The open world action RPG made with the Luminous Engine isn't known to be well-optimized, which shows in our results. There's a vast difference between the average frames per second and the 1% Low FPS.

This is another game powered by the Snowdrop engine, although not quite as impressive as Star Wars Outlaws. Still, even without DLSS Frame Generation (not supported), the RTX 5090 averages 106 FPS with all the bells and whistles turned on and the highest in-game DLSS Super Resolution setting (Ultra Quality).

The latest Assassin's Creed on the market also lacks DLSS Frame Generation, but it's not at all a problem for this GPU, even with DLAA active.

Horizon Zero Dawn Remastered incorporated some of the technical improvements added in Forbidden West. Even with just the old x2 Frame Generation and DLAA enabled, this graphics card averages nearly 200 frames per second.

Turn 10's Forza Motorsport offers a worst-case scenario benchmark where all 24 cars are packed together at the start of a race, with heavy rain pouring. Even so, with DLAA enabled and all ray tracing settings activated (full reflections, ambient occlusion, and the RTGI that was added with a post-launch update), the GeForce RTX 5090 averaged nearly 80 frames per second.

Here's another Unreal Engine 5 game, The Talos Principle 2 by Croteam. This one does exhibit a fair bit of stuttering, as evidenced by the gulf between average and 1% low FPS, though these are game-related issues.

For the final gaming test, I picked Marvel's Guardians of the Galaxy by Eidos Montréal. The game runs on the studio's proprietary Dawn engine. Even with ray-traced reflections and transparent reflections enabled, it performs flawlessly.

NVIDIA also shared an updated build of benchmarking software 3DMark with DLSS 4 Multi Frame Generation, so I ran the DLSS Feature Test, too. This is based on the Port Royal ray tracing benchmark with the addition of DLSS support.

I tested the CNN model at x4 Multi Frame Generation and the new transformer models at x4, x3, and x2; in all cases, DLAA was selected. The CNN model was faster, but just barely.

Overall Thoughts on the NVIDIA GeForce RTX 5090 Founder's Edition

The numbers speak plainly—the new GeForce flagship graphics card practically ate everything I threw at it for breakfast without even breaking a sweat.

It's so powerful that it allows users to push the image quality to the absolute highest level, even in the latest PC games, while retaining very high frame rates. Owners of high refresh rate displays will be delighted to finally find a GPU that can deliver an experience to match 200+Hz screens without having to turn every graphics setting way down.

Of course, much of this is enabled by the enhanced AI capabilities that power DLSS 4 Multi Frame Generation. The role played by Tensor Cores within the new Blackwell architecture has never been more critical, and NVIDIA's choice in this regard is vindicated by the fact that AMD is finally pursuing the same road with AMD FSR 4, which is also going to be AI-based and exclusive to the new Radeon GPUs, just like DLSS is for GeForce RTX cards. AI calculations are going to be increasingly important going forward, and NVIDIA has earned for itself a competitive advantage that's going to be hard to match for competitors. The GeForce RTX 5090 is fully ready for the upcoming RTX Neural Shaders, which could have a big impact on the gaming industry in the future. Once again, NVIDIA is setting itself up for success thanks to a clear vision. Overall, DLSS 4 Multi Frame Generation is yet another big win in all tested games, although I must say I don't see much point in the x3 mode when x4 delivers much higher frame rate boosts with barely any higher latency.

I'm slightly less convinced about the new transformer models. While the Ray Reconstruction one is definitely a win-win already, the Super Resolution model is still in beta and couldn't quite eliminate all ghosting/artifacts in some titles. Moreover, it does require quite a bit of performance in some cases, so I think allowing users the option to manually select CNN over transformer models definitely makes sense, at least for now.

Subjectively, if I had to single out anything of my early experience playing PC games with the GeForce RTX 5090 Founder's Edition, it would be the unprecedented smoothness. The previous RTX generation already enabled high frame rates in most games, of course, but stuttering was almost always there, tarnishing an otherwise great experience. While often a problem caused by game software, NVIDIA acknowledged that the previous Frame Generation method relied on CPU-based frame pacing, which also played a role in those pesky stutters. The new hardware-enhanced Flip Metering largely solves those issues (unless the game has big problems of its own, like Hogwarts Legacy), and it's really a major improvement in how the games actually feel. As a bonus, the graphics card is relatively quiet even under load and definitely quieter - by subjective impression at least - than the RTX 4090.

Granted, all of this assumes you can shell out enough cash to purchase to actually purchase the GPU. Make no mistake: this is as enthusiast as a product can get for PC gamers. I would recommend at least waiting until there's better availability so that retailers will be forced to get closer to the MSRP.

The testing was conducted on a review unit sent by NVIDIA.

Follow Wccftech on Google to get more of our news coverage in your feeds.