• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Fidelity Super Resolution in 2021

JediFragger said:
They've confirmed that no ML is being used though right? In that sense they are extremely limited to what improvements can be made.
Click to expand...

Dunno about confirmed but watching a couple of videos the ultra quality vs native appears to only be spatial upscaling without even any temporal input never mind AI.

Earlier talk suggested it would use a variant of temporal upscaling - which does need information from the game to make work.
 
I don't think AMD is stupid so my view on this is that AMD built FSR real quick just to have something to compete, I believe DLSS really surprised AMD and they were caught unaware by it. They need something ASAP so the current form of FSR was created, doing it using the AI method takes more time to develop.

There is no reason to think that AMD will stop here, they are probably right now working on an AI temporal version of FSR that will come later on and are designing RDNA3 to make use of it - AMD has much more time to get it right for this one, it's still more than 12 months till RDNA3 is announced so they'll spend that time working on a true DLSS competitor/beater

An AI temporal version of FSR may not be as Universally compatible but at least the base FSR upscale fallback is there - when it comes to consoles only the Series X has AI accelerators that could make full use of an AI temporal image reconstruction - and that was because Microsoft asked AMD to give it AI accelerated cores in addition to its base RDNA2 GPU tech. Microsoft had good vision for what was coming so congrats to them and in that regard, it's possible the Series X actually has an RDNA3 feature which gives it a nice advantage
 
Last edited:
i don't think sony and sony playerbase will care about ai accerelates or anything

returnal literally renders at a resolution of 1080p, upscaled to 1440p by temporal reconstruction and then checkerboarded to 4k

now, i don't own the console but every ps5 player out there reports clear and sharp image, either they have no visualized standard for 4k, or it does not matter for these players to render at 1080p or 4k, because current "upscaling" methods seem to do fine for "them"
 
Grim5 said:
An AI temporal version of FSR may not be as Universally compatible but at least the base FSR upscale fallback is there - when it comes to consoles only the Series X has AI accelerators that could make full use of an AI temporal image reconstruction - and that was because Microsoft asked AMD to give it AI accelerated cores in addition to its base RDNA2 GPU tech. Microsoft had good vision for what was coming so congrats to them and in that regard, it's possible the Series X actually has an RDNA3 feature which gives it a nice advantage
Click to expand...
The "AI accelerated cores" are the Big Navi "enhanced" compute units, there is nothing extra in the Xbox. All these rumors are based on an interview with someone from MS where he said they packed a lot of AI in their Xbox ( talking about the ability to do tons of low precision calculations:
"We knew that many inference algorithms need only 8-bit and 4-bit integer positions for weights and the math operations involving those weights comprise the bulk of the performance overhead for those algorithms," says Andrew Goossen. "So we added special hardware support for this specific scenario. The result is that Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations. Note that the weights are integers, so those are TOPS and not TFLOPs. The net result is that Series X offers unparalleled intelligence for machine learning."

Reading this thing the Xbox "supporters" think that Xbox has something extra when in fact he is talking about the RDNA2 "enhanced" CU.
 
I reckon their will be two versions of fsr, one basic for most devices to prove the wide spread of the tech, then their will be a fsr enhanced version which will use rdna2 specifically or maybe 3 i dunno tho id like it to be 2 cos i have rdna2 hehe. I reckon with a slim chance at least if they do a enhanced version it will be for the 6000 series that was initially planned to show off the hardware. How i dunno, maybe use them accelerator rt things to help it with a little bit of ML if thats what they might be able to help with i dunno.

I just dont see them doing only the one they showed at computer show. As it doesnt really show off rdna 2 which i think is what they wanted to do with upscaling tech.

Probably wrong, but who knows, maybe there just getting the basic version out in june 22nd and the better one near end of the year like most thought it was coming out then.
 
SkeeterUK said:
I reckon their will be two versions of fsr, one basic for most devices to prove the wide spread of the tech, then their will be a fsr enhanced version which will use rdna2 specifically or maybe 3 i dunno tho id like it to be 2 cos i have rdna2 hehe. I reckon with a slim chance at least if they do a enhanced version it will be for the 6000 series that was initially planned to show off the hardware. How i dunno, maybe use them accelerator rt things to help it with a little bit of ML if thats what they might be able to help with i dunno.

I just dont see them doing only the one they showed at computer show. As it doesnt really show off rdna 2 which i think is what they wanted to do with upscaling tech.

Probably wrong, but who knows, maybe there just getting the basic version out in june 22nd and the better one near end of the year like most thought it was coming out then.
Click to expand...

Maybe there are reasons why not but I'd like to see the temporal upscaling in Quake 2 RTX get wider adoption - at close to full resolution levels like 80-90% you still get a pretty decent performance increase with almost imperceptible quality decrease and it can dynamically adjust to keep frame rate fluid while minimally impacting perceived quality drop in most cases - even down to 70% which almost doubles frame rate you really have to squint to see the difference which to me is pretty crazy for something that isn't using AI.
 
humbug said:
I don't think we know everything that AMD are doing re: FSR, yet.
Click to expand...

We know what the initial version of FSR aims to achieve and how it will achieve it from https://gpuopen.com/fsr-announce/.
What is AMD FidelityFX Super Resolution?
FSR is our brand new, open source, high-quality solution for producing high resolution frames from lower resolution inputs. It uses a collection of cutting-edge algorithms with a particular emphasis on creating high-quality edges, giving large performance improvements compared to rendering at native resolution directly. FSR enables “practical performance” for costly render operations, such as hardware ray tracing.

How does it work?
FidelityFX Super Resolution is a spatial upscaling technique, which generates a “super resolution” image from every input frame. In other words, it does not rely on history buffers or motion vectors. Neither does it require any per-game training.
Click to expand...

We can speculate on FSR being the framework that AMD will use for a DLSS like competitor with RDNA3, thus supporting titles with be available on launch of RDNA3.
 
Rroff said:
Dunno about confirmed but watching a couple of videos the ultra quality vs native appears to only be spatial upscaling without even any temporal input never mind AI.

Earlier talk suggested it would use a variant of temporal upscaling - which does need information from the game to make work.
Click to expand...


AMD have a patent or 2 on deep learning based super resolution, so it is possible FSR uses DL, but the implementation would be more involved on different hardware.
AMD have neither confirmed nor denied that any DL is going to be used.

AMD told Anandtech that FSR does not use temporal information, and is strictly spatial.


And this is the real crux of the problem, because if neither temporal accumulation nor DL is used, then this is basically a linear scaler like bicubic that has been used by GPUs for decades. Now they do talk about requiring engine integration rather than a driver level post process, so it is possible they get some information form the depth buffer etc, but this won't drastically change the quality that can be achieved.
 
Rroff said:
Maybe there are reasons why not but I'd like to see the temporal upscaling in Quake 2 RTX get wider adoption - at close to full resolution levels like 80-90% you still get a pretty decent performance increase with almost imperceptible quality decrease and it can dynamically adjust to keep frame rate fluid while minimally impacting perceived quality drop in most cases - even down to 70% which almost doubles frame rate you really have to squint to see the difference which to me is pretty crazy for something that isn't using AI.
Click to expand...


UE5's TSR is basically the same. DLSS2 is the combination of essentially the temporal Scaler in quake 2 RTX combined with DL based image reconstruction.
 
oguzsoso said:
Now they do talk about requiring engine integration rather than a driver level post process, so it is possible they get some information form the depth buffer etc, but this won't drastically change the quality that can be achieved.
Click to expand...

The engine integration can simply be the trick they are using to make the FSR widely available. That's why i said it will work on Nvidia anyway, you get the software required through the game engine.
 
Didn't want to start a new thread but DF talks about the limitations of Unreal 5 TSR

Still images look amazing its motion that breaks down quite badly. They think this is aimed at consoles more because you would notice the issues less sitting afar.

 
shankly1985 said:
Still images look amazing its motion that breaks down quite badly. They think this is aimed at consoles more because you would notice the issues less sitting afar.

Click to expand...
That's the nature of TAA it needs more frames and it will get more frames looking the same only when you stand still. When you are moving you need to rely on algorithms and it will break down. Plus you get ghosting most of the time.
The thing is the DLSS has the same problem when in motion but they are only downplaying the TSR because Nvidia has stronger marketing resources. :)
 
shankly1985 said:
Didn't want to start a new thread but DF talks about the limitations of Unreal 5 TSR
Click to expand...

Yeah - I've not been that impressed with Unreal 5 in motion so far - static shots look great but both the Nanite implementation and TSR seem to break down quite badly in motion.

The temporal upscaling implementation in Quake 2 RTX seems to largely avoid these issues while IMO mostly providing superior results to all but DLSS quality mode.
 
Not long to go now, just next 5 days till the 22nd then its out and presumably games patches will be out too to coincide with the launch. Any particular games that people want to use FSR with?
 
You must log in or register to reply here.
Back
Top Bottom