Could be another XeSS like approach, those with dedicated units on newer AMD hardware get the best version and for cross-vendor + legacy compat it's a slightly worse version in both performance and quality. Or a choice for a fallback to FSR 2/3 instead, having a choice between both? The age old image quality or framerate choice would be interesting. Nice to see them confirm its existence though, look forward to seeing it in action and how it improves
Frenetic Pony
Veteran
The quote is just from a handwavy exec telling a team to "do a thing", I wouldn't take it as a quote implying any technical details other than machine learning will be used somehow
An engineer from ARM did an entire paper earlier this year trying for the same goal, seeing if neural networks and temporal upsampling can produce good results at efficient power usage: https://community.arm.com/cfs-file/...iamONeil_2D00_v7_2D00_speakers_5F00_notes.pdf
The conclusion, after numerous experiments, was that a hybrid FSR 2 like framework with machine learning optimized parameters and machine learning taking over some specific functions it was better at like history rectification produced the images with the least failure cases and most efficiency
Another avenue of research has been XESS 2 and it's much more neural network start to end, which Intel has helpfully just given talks about without ever releasing it. "Temporally Stable Real-Time Joint Neural Denoising and Supersampling" is what DLSS 3.5 is based off of, Intel also gave a talk about neural frame generation hole filling that fits in with this same framework. The most interesting frame generation talk in recent years however was predictive frame generation, no waiting for the next frame and then interpolating between, based off history you predict what the next frame is going to be and display that before any new frame is ready, meaning less latency hit (still performance hit obviously).
I wonder which avenue AMD will go with, and if they'll show off a preview of it on their AI presentation next month. Either way they'll almost certainly have joint denoising, they've had a machine learning like (it's just regression based but hey it needs to be fast af) for 2 years now: https://gpuopen.com/download/publications/GPUOpen2022_WALR.pdf
An engineer from ARM did an entire paper earlier this year trying for the same goal, seeing if neural networks and temporal upsampling can produce good results at efficient power usage: https://community.arm.com/cfs-file/...iamONeil_2D00_v7_2D00_speakers_5F00_notes.pdf
The conclusion, after numerous experiments, was that a hybrid FSR 2 like framework with machine learning optimized parameters and machine learning taking over some specific functions it was better at like history rectification produced the images with the least failure cases and most efficiency
Another avenue of research has been XESS 2 and it's much more neural network start to end, which Intel has helpfully just given talks about without ever releasing it. "Temporally Stable Real-Time Joint Neural Denoising and Supersampling" is what DLSS 3.5 is based off of, Intel also gave a talk about neural frame generation hole filling that fits in with this same framework. The most interesting frame generation talk in recent years however was predictive frame generation, no waiting for the next frame and then interpolating between, based off history you predict what the next frame is going to be and display that before any new frame is ready, meaning less latency hit (still performance hit obviously).
I wonder which avenue AMD will go with, and if they'll show off a preview of it on their AI presentation next month. Either way they'll almost certainly have joint denoising, they've had a machine learning like (it's just regression based but hey it needs to be fast af) for 2 years now: https://gpuopen.com/download/publications/GPUOpen2022_WALR.pdf
Last edited:
D
Deleted member 2197
Guest
I'm not sure devs would want to implement both FSR 4 and FSR 2/3 in new games considering the development time involved. Not even sure how many will be implementing FSR 4 into existing FSR 2/3 games. Is it left up to the studio and whether their maintenance budgets include adding new gpu features or do IHVs support these projects via onsite support and financially?
DegustatoR
Legend
My guess would be that AMD will make use of DirectSR and provide either option through their drivers.
Still need games to use DirectSR though. Wonder who will pull the trigger first.
DegustatoR
Legend
It's still not released so unless it will be no games will use it.
raytracingfan
Newcomer
Ideally, the FSR 4 SDK will have the same API and inputs as previous FSR versions and include both the AI version and the previous version. If this is the case developers will only implement FSR 4 in future games to support both AI and non-AI FSR and FSR 4 would be a drop-in replacement for existing FSR integrations.
DavidGraham
Veteran
AMD is saying that Black Ops 6 will integrate the next generation machine learning based FSR soon.
oh so very interesting.
this would imply that it would just run off their CUs as UDNA is not here yet.
The real question is what feature sets are required (if any) to support this (as typically FSR has very low feature set requirements), and secondly, how much compute is needed for it.
Considering how well BLOPS6 is doing on the market, good to attach yourself to it while it's hot. I don't know when it's coming, but I also hope it's soonTM.
lol, can't help but feel you are a LucasArts adventure player. Whenever I see anything close to SCUMMM type text, it always gets me.
