Just as it leads in data center AI accelerators, NVIDIA’s lead in the PC graphics space has been dominant for years and generations of products. When the company announced its forthcoming GeForce RTX 50 series of graphic cards based on its Blackwell GPU architecture back at CES in Las Vegas this year, there was almost no question as to whether NVIDIA would have the most powerful product on the market, competitively, when cards ship to gamers later this month. However, details and the nuanced picture of performance with the new GeForce RTX 5090, which is first out of the gate, had to wait until today’s embargo lift.
So, without further ado, say hello to NVIDIA’s new GeForce RTX 5090 Founders Edition graphics card for PC gamers, enthusiasts, and creators, and then let’s unpack what’s on tap and what to expect from it.
Under The GeForce RTX 5090’s Hood: NVIDIA’s RTX Blackwell Architecture
We won’t wade too deep into the weeds here, but it’s safe to say that Nvidia built its RTX Blackwell GPU architecture from the ground-up for enabling advanced rendering technologies, like Deep Learning Super Sampling with frame generation and neural shaders, which employ AI for improved shader visuals and performance. In addition, Blackwell’s 4th gen ray tracing cores are also built for handling clusters of mega geometry and standard geometry together with better fidelity and efficiency, while its 5th gen Tensor cores now support FP4 precision for up to double the throughput for DLSS 4 and AI workloads in general.
Shader Execution Reordering has been improved in Blackwell by up to 2X, and Nvidia’s new GPU architecture has also been bolstered with a new memory controller that, when coupled with 32GB of GDDR7 memory, offers twice the bandwidth of GDDR6, at up to 1.8TBps (for RTX 5090) with better power efficiency as well. Finally, and not that my short list here is exhaustive, but different blocks of Nvidia’s RTX Blackwell architecture can enter deeper sleep states and wake from those sleep states much faster than the company’s previous gen Ada architecture, with advanced clock gating, power gating and voltage rail gating. In fact, Blackwell’s entire clock tree can be disabled while the GPU is still active, so if other parts of the board like memory are idle, additional power can be saved. Again, these are just a few of the advancements in RTX Blackwell, but if you want an even deeper dive, head on over to my colleague Marco’s coverage at HotHardware for the full download.
AI Enhanced And Generated Frames Are The Future, So Get On Board
There’s been a lot of chatter in the gaming community on this topic, so let’s address it head on. Nvidia made some bold performance claims for its GeForce RTX 50 series GPUs, with CEO Jensen Huang pitting its new $549 GeForce RTX 5070 against the company’s $1599 GeForce RTX 4090 previous generation flagship, noting they offer similar performance. The caveat here is that the new RTX 5070 will deliver that class of performance via advanced technologies that are enabled not only in its chip architecture, but with new AI rendering techniques like DLSS 4 with multi-frame generation.
In a nutshell, Nvidia’s previous generation DLSS 3 tech could render a single AI-generated frame in a game engine, but now with DLSS 4, the neural network can generate up to 3 additional frames in between traditionally rendered ground truth frames. Along with advanced display engine frame pacing onboard Blackwell, this affords a massive performance lift for the DLSS 4 capable GeForce RTX 50 series, versus standard single-frame generation on board legacy GeForce RTX 40 hardware. Of course, purists are going to argue about “fake frames” and the inherent latency with interacting with that many AI-generated frames in gameplay. And though Nvidia’s new Reflex 2 technology is targeted at mitigating frame-to-reaction latency, game developers have to provide support for both DLSS 4 and Reflex 2 in their game engines.
At launch, Nvidia was quick to point out that there are 75 games on the market that will support DLSS 4 with multi-frame gen, and a that list will grow with time, as it did for the hundreds of titles that now support previous versions of DLSS. However, again, some negative purists in the community will argue that there’s no replacement for brute force horsepower in traditional rendering. This is true at face value, but the problem is, as game engines with ray tracing and other advanced rendering effects get more sophisticated and compute-intensive, throwing silicon at the problem, without advancing rendering techniques with AI and other algorithmic approaches, quickly reaches a point of diminishing returns.
In fact, in conversation with all of the major players in PC graphics, AMD, NVIDIA and Intel, all point to an AI-assisted frame generation future as the way forward. So, it’s not a matter of if, but when the game developers and the gaming community at large get on board.
A High-Level Performance Profile Of GeForce RTX 5090
Breaking down performance of Nvidia’s new GeForce RTX 5090 can be carved up fairly cleanly from three high level angles — traditional raster performance, performance with super-resolution and frame generation, and of course for non-gaming applications, AI inferencing and generative AI performance, as well as content creation.
Speaking anecdotally with respect to content creation, my experience with GeForce RTX 5090 in video rendering was met with a significant reduction in transcode times for my final renders. This is due to the GeForce RTX 5090’s additional encoder on board; where RTX 4090 has two encoders, RTX Blackwell 5090 has three that support Nvidia’s 9th generation NVENC. And with that covered, let’s look at a quick sanity check on AI throughput.
ML Commons recently released a PC client version of its widely respect MLPerf AI benchmark. Here we see the GeForce RTX 5090 outpace its previous generation 4090 sibling by over 38%. I’ll also note that, in addition to this token throughput gain, time to first token latency has been reduced by about 25% as well.
Next, let’s dig into cutting-edge gaming performance, starting with the wildly popular action RPG title, Cyberpunk 2077 at its Ultra image quality preset.
Here we see the full might of GeForce RTX 5090 Blackwell at play, with the top long bar representing its fastest DLSS 4 rendered frame rate with a 4X frame generation setting in the game. This setting actually renders three AI generated frames, along with a single upscaled frame, in order to maximize performance. If we compare the fastest setting for RTX 4090, which is DLSS with single AI frame generation, we see over a 65% performance lift for RTX 5090 comparably, and nearly 3X the performance of a previous gen GeForce RTX 4090 with no frame generation turned on. Looking at performance with frame generation turned off completely, RTX 5090 is over 22% faster than 4090. And AMD’s fastest Radeon GPU currently isn’t even in the hunt with FSR 3 super resolution enabled (sans frame generation, which is not available for Radeon in this game title).
The Formula 1 racing sim F1 24 shows a tighter spread in performance for GeForce RTX 5090 versus the previous generation RTX 4090, with about a 37% lift when frame generation is enabled, and a 24% edge with no frame AI frame generation. This game title also shows AMD’s best foot forward for the Radeon RX 7900 XTX, since F1 24 does support AMD Fluid Motion Frames frame generation tech, enabling the Radeon to move up the stack significantly, though it is no match for an RTX 5090.
Finishing-up again with a modern, gorgeous and graphically rich game title, Black Myth: Wukong definitely puts a strain on all GPU contenders at its Cinematic quality preset with full ray tracing employed. Here the new GeForce RTX 5090 is about 33% faster than RTx 4090 at similar settings with frame generation turned on, while if you flip frame gen off, the spread drops to about a 28% advantage for RTX 5090. And once again, even with frame generation enabled, the fastest Radeon card on the market right now can’t even hit playable frame rates at these settings in this game title.
By the way, make sure you swing on over to HotHardware for their full copious gauntlet of benchmarks with GeForce RTx 5090, to fully complete the picture.
Key Take-Aways And Future Projections For NVIDIA’s GeForce RTX 50 Launch
It’s safe to say there has been a bit of controversy over Nvidia’s new GeForce RTX 50 series, but more so with respect to pricing, which increased 25 percent, at $1999 for the RTX 5090 versus the GeForce RTX 4090 which retailed for $1599 when it launched. There’s little question that a $2,000 graphics card is just too expensive for a lot of gamers, but when you’ve got easily the most powerful, feature-rich product on the market, and it took a serious engineering investment to develop it, you can command a price for that product that your major competitors simply cannot. It’s that simple.
Then again, if all you’re focused on is the price of a GeForce RTX 5090 Founders Edition, I think you’re missing the point. Nvidia’s engineering investments for GeForce RTX Blackwell will also pay off at lower, more mainstream price points as well.
And that ultimately is where the rubber meets the road. Folks in the community will debate the merits of Nvidia’s pricey new top dog graphics card, but you can’t argue the merits of the advanced rendering technologies that RTX Blackwell brings, and the performance and visual fidelity they enable. When Nvidia CEO Jensen Huang noted the company’s forthcoming $549 GeForce RTX 5070 will be as fast as its previous generation top-end GeForce RTX 4090, it was with specific intention to illustrate the merits of the AI-assisted rendering techniques the company’s has delivered with this family of PC graphics cards, developer tools and software.
Advancing PC graphics via AI-generated content is a natural evolution, and the way the industry is moving as a whole, not just Nvidia. However, once again, Nvidia is also on the forefront of GPU architectures, along with the engineering and software support required to blaze this new trail of innovation. The new GeForce RTX 5090 is the fastest PC graphics card money can buy currently, and it asserts the company’s dominance in a product category and technology that it founded itself on so many years ago.
