When Nvidia released their RTX 20 series line of GPUS they also announced two exciting new technological improvements: their GPUs would form then on be able to perform real time and would feature Deep Learning Super Sampling (DLSS). We’ve already explained what ray tracing is in a separate article, but what exactly is this DLSS technology?
Simply put, it’s an AI technology that learns from past experience in order to help your game run smoother.
Using AI to improve GPU performance and quality
DLSS is a piece of technology that uses AI (Artificial Intelligence) to learn from experience in order to ‘predict the future’. This sounds pretty vague and futuristic, so let’s start with an example.
Say you’re looking through old photos of yourself on the internet. Back then, photos were stored online in a very low resolution (due to a variety of reasons which aren’t important to this example) so online photos from years ago will be rather pixelated and blurry. When you’re looking at those photos you know what it’s supposed to look like though, so you’re able to ‘look past’ the subpar quality of the photo and you know what the scene pictured in the photograph is supposed to look like. You know that your friend’s face doesn’t have blocky, jagged edges, so you can basically ignore that and ‘fill out’ the picture yourself.
DLSS is a technology that’s trying to replicate that, in essence. It is an artificial intelligence network which does trillions of operations through a supercomputer that’s called DGX-2 (which has 16 GPUs in it) in order to study and learn what ‘the perfect frame’ looks like. The perfect frame, in this case, is ‘Ground Truth’, which is basically a reference image indicating how the rendered image should look like in an ideal world.
DLSS is constantly learning, in an ever-going back-and-forth between the rendered image and the Ground Truth image with the DGX-2 supercomputers, how to more closely approach the perfect frame. DLSS gets fed all sorts of different images as well; it’s learning how to render cartoon-style graphics, photo realistic images, picturesque outdoorsy scenes, and so on.
The idea with this that in the (near) future, the GPU will be able to render the original image at a slightly lower resolution, run it through DLSS, and then put out a higher resolution image. Because of the fact that the GPU doesn’t have to do all that work rendering the higher resolution images it will also increase the performance. ‘The dream,’ according to NVIDIA, is to have (for example) a 10 Teraflops GPU perform like a 30 or even 40 Teraflops GPU.
To put it shortly and very simply: DLSS an AI technology which teaches your GPU what certain images should look like in a perfect scenario, so that it doesn’t have to do all the ‘thinking’ on the spot.
You can learn more here.