As a quick refresher, the NVIDIA GeForce RTX 5090 sits on the top of NVIDIA’s consumer, gaming focused lineup of graphics cards. Based on the new Blackwell architecture which is being utilized across the datacenter AND gaming offerings from NVIDIA this generation, the RTX 5090 is powered by the GB202 die which is built on TSMC’s N4P process.
While the RTX 5090’s GB202-based GPU contains a staggering 21,760 CUDA cores, it’s the improved AI acceleration features, and more importantly what the future of AI acceleration in gaming looks like, that NVIDIA is pushing for this generation. From a specification perspective, the RTX 5090 alongside its 32GB of GDDR7 memory is capable of 3352 AI TOPS, compared to 1321 TOPS for the RTX 4090 that precedes it.
With the RTX 5090’s improvement in traditional rasterized rendering over the RTX 4090 also comes a whole host of new features aimed at using modern AI acceleration techniques such as transformers to dramatically increase frame rate, as well as decreasing graphics memory consumption.
NVIDIA’s DLSS AI-based upscaling and frame generation technology sees a big overhaul alongside the launch of the RTX 50-series GPUs. First, NVIDIA has migrated the DLSS upscaling model from their previous convolutional neural network (CNN) model to a more modern transformer-based architecture providing increased image quality and improved temporal stability.
The second part of the DLSS updates with Blackwell come in the form of what NVIDIA introduced with the 40-series, frame generation. Previously, frame generation was limited to generating one frame per input native frame. Now, the DLSS 4 model is capable of scaling to 3 or 4 frames, depending on whatever setting the user chooses.
Games that were previously released with DLSS 2 and 3 support won’t be left behind though, as NVIDIA is providing a user accessible toggle to change from the CNN to transformer model and from 2x to 3x and 4x frame generation where applicable in older titles which might not be updated with the new DLSS 4 SDK.
NVIDIA’s new RTX Neural Shaders are part of an SDK that enables developers to leverage deep learning inferencing of their game’s assets such as textures and materials to provide memory compression and increased performance while maintaining or even improving visual quality. So far NVIDIA has developed a tech demo of these features, but there is no implementation in actual shipping games yet.
While Neural-based rendering, and particularly the improvements brought with DLSS 4 have a ton of potential to rethink how we measure gaming performance, first let’s look at what the RTX 5090 offers compared to the previous leader of the pack, the GeForce RTX 4090, released in October 2022.
When we add AI-based upscaling (using the new transformer model of DLSS 4) and first-generation frame generation (part of DLSS 3), we see massive improvements to the average frame rate. Take Indiana Jones for example; while frame rates where just 29 FPS or 38 FPS for the RTX 4090 and RTX 5090 respectively, with the upscaling and frame gen enabled, that increases by about 3.2x, reaching above 120 FPS on the new RTX 5090 GPU. Hogwarts jumps from 87 FPS to 230 FPS on the RTX 5090. These are huge improvements in frame rate and result in a generational life on like-for-like features vs the RTX 4090 on the order of 30%.
If we dig deeper into the titles that we have tested which have full DLSS 4 implementations, allowing us to choose the level of frame generation in game (2, 3 or 4 frames), we see some incredibly high scaling in average FPS.
Hogwarts Legacy is an interesting case for DLSS 4 frame generation because despite having all the visual quality features turned on, the RTX 5090 is still capable of delivering a quite respectable 87 FPS at 4K native. Here, it might make sense to forgo the upscaling portions of DLSS which may have an impact on visual quality, and instead to just enable the frame generation parts of DLSS 4, resulting in a 264 FPS average FPS. But the maximum frame rate of 392 FPS is 4.5x faster than the full native frame rate of 87 FPS.
Cyberpunk 2077 shows similar scaling in average FPS with all the DLSS features enabled but has a slightly different story. When the native frame rate is low (in this case 34), DLSS Frame generation alone will not provide a great gameplay experience due to input latency. Since the game engine is still sampling user input based on the native frame rate, despite having up to 119 FPS with DLSS 4 enabled and no upscaling, the user’s experience will feel similar to the native 34 FPS situation. In this case, upscaling plus frame generation greatly decreases the input latency, to around a 95 FPS experience in this case.
Just looking at the few quick gaming results that we have today, it’s clear that the NVIDIA GeForce RTX 5090 is a force to be reckoned with in PC gaming. Given that the RTX 4090 was already the fastest gaming GPU in the market, the additional improvements for the RTX 5090 mean that NVIDIA continues to leave competitors like AMD and Intel further behind.
Looking beyond gaming performance, there are a few non-gaming AI-based benchmarks on the PC platform that have started to come into focus for measuring real-world performance of AI operations on client PCs.
UL Procyon now features a suite of AI tests including more traditional AI computer vision models and generative AI of both image and text using transformer-based models such as Stable Diffusion and Llama 3.x.
Going from the RTX 4090 with its 512 4th generation tensor cores to the RTX 5090 with its 768 5th generation tensor cores provides an average increase of 29% in LLM performance as measured by UL Procyon.
NVIDIA and UL also provided a preview version of the UL Procyon AI Image Generation benchmark with a cutting-edge FLUX.1 Dev model from Black Forest Labs. Here we can see in FP8, the RTX 5090 is able to generate images 75% faster than the RTX 5090. When moving to FP4, which is natively supported by the new 5th generation tensor cores, the speedup goes to over 4x when compared to RTX 4090 and 63% compared to the RTX 5090 in FP8 mode.
One thing that the RTX 5090, at least the NVIDIA-produced Founders Edition card that we are looking at today, that didn’t increase over the RTX 4090 is the physical size.
Despite most retail RTX 4090 graphics cards occupying over 3 PCIe slots, the RTX 5090 Founders Edition is a more traditional dual slot form factor. NVIDIA was able to shrink down the width of the RTX 5090 using a dual fan, completely flow-through cooling design.
On the RTX 5090 Founders Edition, the main PCB of the graphics card is quite small compared to traditional high-end graphics cards and sits in the middle of the cooler. Connected to the GPU is a 3D vapor chamber, which has heat pipes extending to both sides of the card, allowing the two cooling fans to flow air across the heat pipes, unrestricted by other components.
There have been some initial concerns about the community about this cooler, and its potential inability to provide adequate cooling for the massive 575W power draw of the RTX 5090. But so far in our testing, we have found no concerns with the cooling design of the RTX 5090 Founders Edition.
Compared to our ASUS RTX 4090 TUF Gaming OC, which is a 3+ slot card with 3 fans, the RTX 5090 holds its own, reaching a peak temp of 75C compared to 71C on the RTX 4090.
Despite the small temperature difference between the two cards, the RTX 5090 Founders Edition consumes an average of 550W, compared to 431W for the RTX 4090 TUF Gaming OC card.
Additionally, we see no observable thermal throttling from the RTX 5090 Founders Edition over time.
This testing was conducted on an open-test bench, but we believe that any system that already has adequate airflow for a high-end graphics card should be sufficient for the RTX 5090 Founders Edition.
The class leading gaming performance, AI performance, and small form factor cooler design do come at a cost though. The NVIDIA GeForce RTX 5090 Founders Edition will be offered for sale on January 30th for $1999 USD. Other GeForce RTX 5090 offerings will be available from partners such as ASUS, MSI, and Gigabyte with from $1999 and up.
Research commissioned by: