DLSS 3 Frame Generation: Double the work for the CPU? Surprising results!

The graphics card calculates with DLSS performance a quarter-resolution image, then this is held back as the first frame. When the second frame, also created in reduced resolution using DLSS, has been calculated by the graphics card, the first frame that has been retained so far is used to interpolate a completely new frame created by the AI ​​from the information in the first frame using the second frame. Sounds complicated, but it’s actually quite simple. Frame 1, the generated AI frame (Nvidia calls this “Frame 1.5”), is then output before the second frame. The latter also serves as the basis for the following AI calculation (“Frame 2.5”). So DLSS 3 always renders a frame, followed by an AI frame, then another calculated frame, then another AI frame. So every other frame with DLSS 3.0 is a “fake” image generated by the AI, which is inserted between actually rendered frames, from which information the AI ​​frame was created. In this case, the graphics card only renders an eighth of the pixels actually visible on the screen. The rest comes from the upsampling and the alternately inserted AI frames. This article is only about the topic of frame generation in the CPU limit, you will find the graphic effects, including on performance, in the large practical test for DLSS 3.

DLSS 3 Frame Generation: double the work for the processor? Surprising results

Source: Nvidia

However, what Nvidia didn’t mention at all or hardly noticed during the presentation is the processor. Not the graphics processor, but the central processing unit, also called the CPU. When you play with the computer, the image is always displayed on your screen at so many frames per second. Many play with 60 fps, some 30 fps are enough and others like to play with over 200 fps – that’s a matter of taste and depends on your own requirements. However, it is always the case that these images are calculated by the graphics card and processor. The latter prepares the images for the graphics card, which then outputs them to the screen. For example, if you are running at 80 fps, then both the GPU and the CPU are “burdened” with these 80 fps. If the user now installs an RTX 4090 and activates DLSS 3 including frame generation, the 80 fps can quickly become 160 fps. The question, which this article is supposed to clarify now, is whether these additional frames are noticeable in the processor utilization or not. Because logically it should be like this – the additional fps that DLSS 2 has freed up have so far always led to an increased CPU load if no frame limiter was active.

DLSS 3 Frame Generation: The test setup

From the start, we didn’t intend to test this feature with 20 or 30 different processors. The reason is simple: there are very few games that support DLSS 3 and the function is still very new and only runs with beta drivers. We would like to give the whole thing a little more maturing time so that we don’t blindly run into teething problems and have to recreate our measurements over and over again. We therefore only used two processors: a Core i7-8700K, representative of older gaming computers, and a fairly recent Core i5-12600K, which represents an ordinary, high-priced gaming PC. In both cases, of course, an RTX 4090 is used as well as 32 GiByte RAM with manufacturer clocking. As a benchmark we use F1 22, with an update for a beta build, which enables DLSS 3 and frame generation. The details have been maximized including ray tracing, the resolution is UHD and DLSS performance is active. The tests were carried out on the regular CPU test track since we want to determine the effects on the processor.

DLSS 3 frame generation: the benchmarks

The increase of over 90 percent can hardly be put into words and is very impressive. The big surprise, however, is that these higher numbers hardly to not at all noticeable in the CPU utilization. In the case of the Core i5-12600K, the average load even drops by one percentage point. We also observe this behavior with the smaller Core i7-8700K, which is now no longer as powerful for modern games – the load fluctuates only minimally here as well and bears no relation to the almost double fps. If the CPU actually had to calculate the displayed fps, the displayed load would be much higher on the one hand and the processor’s power consumption on the other. However, the difference between the two benchmarks is so minimal that one can speak of measurement inaccuracies (a benchmark never runs 100 percent exactly the same).

The frame times are dynamic, please click on a CPU to show or hide it. Above all, the frame times are positive, especially with the Alder Lake processor. An i7-8700K, on ​​the other hand, is apparently not powerful enough to smooth out the peaks sufficiently quickly. Despite everything, the Coffee Lake CPU also sees a noticeable improvement. Overall, there are fewer spikes and lower frame times mean a smoother picture.