In a brief gameplay demo with Cyberpunk 2077 & DLSS 3, NVIDIA’s top-tier GeForce RTX 4090 graphics card displayed some impressive numbers. The NVIDIA GeForce RTX 4090 Founders Edition graphics card, which we thoroughly reviewed here, was used in the demonstration. Using Ultra Ray Tracing and Psycho settings, the card was running Cyberpunk 2077 at 1440p when connected with an Intel Core i9-12900K computer (or maximum settings). Without DLSS and Reflex, the game operated at its native resolution at an average frame rate of 59 FPS (49 FPS 1% Lows), with a latency of 72–75 ms.
The GPU was operating at over 2.8 GHz at these settings, averaging approximately 2810-2850 MHz (min/max), while at 100% usage, the temps remained constant at 50 to 55 C. This represents an increase of 13% (up to 330 MHz) over the reference boost clock of 2520 MHz, and what’s really astounding is that no overclocking was done.
All of this was going on at stock. There are other games that provide even faster clock rates than this one. Since then, it has been claimed that the card can be overclocked to operate at above 3 GHz.
Performance and settings will differ in the final version of the game because it was using a pre-release version of DLSS 3. The DLSS Frame Generation setting becomes active as soon as the DLSS 3 setting is turned on. We once again saw full GPU use with over 2.8 GHz rates when using the “Quality” option, however this time the temperatures were more in the range of 50C than 55C. (keep this in mind).
With DLSS 3 enabled and an average latency of 53.00ms, the NVIDIA GeForce RTX 4090’s performance increased to 170 frames per second (FPS) on average (119 FPS on 1% Lows).
That represents a 2x increase in FPS and a 30% decrease in latency compared to DLSS deactivated. NVIDIA also published the wattage statistics with both DLSS deactivated and enabled using the most recent PCAT tool, which supports the PCIe 5.0 16-Pin (12VPHWR) power connection. The NVIDIA GeForce RTX 4090 graphics card used an average of 461 Watts with DLSS 3 deactivated and had a performance per watt (Frames/Joule) of 0.135 points. When DLSS 3 was activated, the GPU’s power decreased to 348W, a 25% decrease. Additionally, this resulted in a 0.513 perf per watt gain, a 3.8x increase.
The fact that the DLSS algorithms are being executed on tensor cores rather than the FP32 cores may be one of the factors contributing to the power numbers being such a large concern. At certain points, the cores are specialised, and rather than brute-forcing the entire GPU, which leads in a larger power draw, the tensor cores can handle the data much more quickly and efficiently while resulting in lower power usage.