A new async approach saves lots of performances with DLSS5 on AMD, could work on Nvidia too
Hi everyone, i just released my last "experiment". a new implementation of DLSS5 on AMD, completely open source.
But what's it's even more interesting, i managed (with a huge help of Claude, tbh) to achieve an asyncronous system which is able to render the dlss effect without waiting for each frame.
For example, i tested it on an RDNA3 card, at 1440p with raytracing disabled and FSR4 + frame gen. on this configuration, the game normally can reach a pretty high framerate, in this example it's almost at 139fps. The dlss5 network runs very slowly there, and waiting for it at each frame would mean running at around 20 fps at 1440p, even if applying dlss5 before the upscaler.
With this new approach, the dlss network is applied only "every now and then" (as soon the dlss5 network finishes working on the previous frame), and every frame is composed based on the previous frames. when the dlss5 network finishes rendering, that frame is already "old", so the software takes a residual (network output - network input) and applies it to the current frame using a displacement map.
Performance-wise, this means that the final framerate will be still pretty close to the original. in this example, 119 fps vs 139.
The drawback: On the current state, it can introduce some artifacts, wich aren't an issue when the game is still (and that's why it isn't noticeable in the screenshots), but it can be noticed in some situations with movement. It's usually more noticeable the higher is the difference between the network's and the game's rendering times. This also means that frame gen makes this even worse.
It also differs from game to game, and can be more or less noticeable in some scenes than in others.
But i'm pretty sure this can be worked on, at least in part. i already managed to reduce it a lot before the first release.
Plus, i also managed to make it work on Linux/Proton.