AMD could be planning to take a different direction with FSR in the future, and if a clue highlighted on Twitter is anything to go by, Team Red’s intention is to develop the frame rate boosting tech along AI-powered lines – just like Nvidia DLSS.
That theory is based on a hint tweeted by well-known hardware leaker Greymon55, who spotted a new commit in the LLVM repository which is about the introduction of WMMA (Wave Matrix Multi-Accumulate) instructions on GFX11.
AI accelerate! Should work with FSR3.0 and other features. https://t.co/Wh31UxdoaIJune 29, 2022
Okay, that sounds like a load of gobbledygook, so how about a translation? Essentially, GFX11 refers to AMD’s incoming next-gen RDNA 3 GPUs (and Radeon Pro cards), with WMMA instructions being a way to really boost machine learning (AI) operations. Therefore this could point to FSR – perhaps in version 3.0 – making use of such AI chops for better quality upscaling results.
One theory floated off the back of this is that perhaps RDNA 3 graphics cards will offer some kind of built-in hardware functionality along the lines of Nvidia’s tensor cores – dedicated AI processors on RTX graphics cards, which are used as muscle to drive DLSS.
Analysis: AMD to take the next turn-off into AI avenue?
Obviously, we need to be very careful about leaping to big conclusions from a simple commit. And certainly this does not mean that next-gen RDNA 3 graphics cards will necessarily come with specific hardware designed to accelerate AI workloads (like Nvidia’s tensor cores).
There are a few key things to consider here. Let’s not forget that AMD has only just launched FSR 2.0, and with that tech, a lot of fresh moves were made including the transition to temporal upscaling, offering big improvements over spatial upscaling (as used by FSR 1.0). We’ve seen that this makes a major difference in supported games, and comparisons indicate some pretty impressive image quality results for FSR 2.0 compared to DLSS 2.0, with Team Red catching up in a big way.
Furthermore, AMD has argued that FSR 2.0 is an effective rival for DLSS without the need for any AI-powered acceleration (or indeed dedicated hardware on the GPU to improve those operations – which takes up space that could be used for pushing other aspects of GPU performance). Indeed, recently AMD very much played down the influence and importance of AI in upscaling technologies…
So, for AMD to be looking at equipping its RX 7000 graphics cards with some kind of equivalent for tensor cores seems a fair old stretch, particularly as we’ve not heard anything from the rumor mill about this before now – and these next-gen GPUs are not very far off launching (they’re possibly arriving in October).
Certainly, RDNA 3 cards may make use of those WMMA instructions in some way to further optimize FSR 2.0, perhaps, but any implementation of FSR 3.0 using machine learning and dedicated hardware married up with that is surely a good way down the line. What this does seem to point towards is that AI is indeed the road AMD is looking down for the future (despite those recent comments and doubt cast on exactly how much difference it really makes for DLSS).
In that case, the plan may be to keep FSR 2.0 in use going forward after the next-gen implementation (FSR 3.0) arrives, just as with FSR 2.0 debuting, games are still getting FSR 1.0 support (as the latter may not be as good, but offers a wider range of coverage for lesser-spec GPUs).
Who knows, ultimately, but what is clear enough is that AMD has built and pushed FSR on being an open standard, and usable across not just its own Radeon graphics cards (even somewhat older models), but also Nvidia’s and Intel’s GPUs. So whatever happens, presumably AMD won’t want to turn its back on that central piece of philosophy for its frame rate booster in the future.