Just when AMD and Intel thought they had a chance.
Audio player loading…
At GTC 2022 Nvidia unveiled the next step for its Deep Learning Super Sampling technology and it promises huge improvements to frame rates thanks to a complete rethinking of how it works. In fact, Nvidia claims its new DLSS 3 is capable of quadrupling frame rates compared to native. This is on top of the new GPUs offering 2–3 times the ray tracing performance (opens in new tab) of the previous generation.
Nvidia showed off its new DLSS 3 tech in Cyberpunk 2077, which saw the frame rate shoot up from 22fps at native 4K up to 85fps with DLSS 3 using a new Ray Tracing: Overdrive Mode. That’s a tasty improvement for sure, although Nvidia promises even higher improvements in Portal with RTX, which saw a 550% increase in performance according to a graph on its updated DLSS mini-site (opens in new tab).
The first iterations of DLSS worked through a combination of a low-resolution current frame and a high-resolution previous frame to build the high-resolution current frame on a pixel-by-pixel basis. With DLSS 3, the new 40-series GPUs are capable of producing entirely new frames rather than just pixels.
There are four components to this new way of working though, and that means you’re going to need either an RTX 4090 (opens in new tab) or one of the new RTX 4080s in order to access this new DLSS technology. Core to this is a new Optical Flow Accelerator, which works with the game engine motion vectors, which show how pixels are changing, the convolutional AI frame generator and its Reflex super low latency pipeline to build those frames.
It processes the current frame and the previous frame to see how the pixels are changing to produce entire new frames. And it does so without processing the graphics pipeline, and that means it can benefit games that are both GPU-bound and CPU-bound.
(Image credit: Nvidia)
As can be seen from the graph above, DLSS 3 can have a massive impact to games. We’ll have to see how this plays out when we get the new cards in our test rigs, but these are promising numbers at the very least. We’ll have to see if there’s any lag with creating several frames at once, but given Reflex gets a mention in the creation of the frames, we’d assume this is covered.
As Jen’Hsun succinctly put it when finishing off the segment on DLSS 3, “DLSS is one of our greatest neural inventions.” And from the improvements shown in the games at GTC, he looks pretty spot on.