Fidelity Super Resolution in 2021

[nvidiamd]

Alex on Control DLSS 1.9:
"Yeah it looks a lot better than native res in this game and is pretty cheap"

Alex on Control DLSS 2.0:
"Of course, this isn't the first DLSS implementation we've seen in Control. The game shipped with a decent enough rendition of the technology that didn't actually use the machine learning Tensor core component of the Nvidia Turing architecture, relying on the standard CUDA cores instead. It still provided a huge performance boost, and generally looked better the higher up the resolution chain you went, but the new 2.0 revision offers a profound improvement.Temporal ghosting is massively cut back, while break-up on sub-pixel detail and transparent textures are reduced to a minimum. Impressively, we were able to find examples of the new DLSS 2.0 at 1080p delivering improved visual quality over the older version running at 1440p"

 

[Rroff]

You can post some pictures from Quake II native vs TAAU with RT reflections and we will see how good they are upscaled.

Having trouble with imgsli at the moment for some reason so will have to post the individual pictures:

Spoiler

EDIT: Frame rate is relatively low because I'm running it modified with 8 reflection/refraction bounce passes and 8 light bounces vs stock 2 and 1.

 

[Godhand007]

Well, we are saying that DF has sold out, which is true but as a business decision it's not a bad thing. Create credibility for few years and then get a large payout, remember Tom's Hardware. Who cares what happens later on. People like Richard from DF would be retiring soon (I am just guessing). It's people like Alex who will have to live to with this.

 

[Rroff]

Spoiler

Not bad for a 1997 game though I kind of ran out of artistic talent before the engine did.

The path tracer is perfectly capable of rendering a whole scene to that kind of quality/detail level as the spent casings (or better if you have actual artistic ability) though.

 

[Godhand007]

The tech is less important than the people. The difference between FSR UQ and DLSS is small, the difference in user base between 15% and 100% isn't.

Also, most people like cheaper things with good enough quality. Not everyone is interested in funneling money for shareholders and profit gains of a company, of which, only a limited amount is used for R & D.

 

[Godhand007]

Spoiler

Not bad for a 1997 game though I kind of ran out of artistic talent before the engine did.

The path tracer is perfectly capable of rendering a whole scene to that kind of quality/detail level as the spent casings (or better if you have actual artistic ability) though.

Makeup on a pig situation with those textures though. Nvidia could have spent that money and resources on supporting a new game/studio.

 

[Rroff]

Makeup on a pig situation with those textures though. Nvidia could have spent that money and resources on supporting a new game/studio.

More the underlying geometry than the textures - I'm zoomed in quite a bit on the textures - they didn't really spend a lot of money on it though someone already had a path tracer implementation just lacking the quality/performance via doing it on RT capable hardware and a lot of the rest of the assets like the weapon models are pulled from Q2 Pro, etc.

It is a shame they didn't have someone go in implement some kind of static mesh type system, decals, redo the monsters in high resoltuion and build a couple of demo levels from scratch with ray tracing in mind though to show it off.

 

[nvidiamd]

Having trouble with imgsli at the moment for some reason so will have to post the individual pictures:

Spoiler

EDIT: Frame rate is relatively low because I'm running it modified with 8 reflection/refraction bounce passes and 8 light bounces vs stock 2 and 1.

If you look at the wooden frame in the 1st set of pictures the degradation is visible, look at the middle of each window at the wooden frame. not too bad tbh but if you put them on slide it is there. Also the shadows are softer.
Even so running with higher RT than you usually get in games, will ruin the experiment, it is like upscaling from 4k to 8k vs 1440p or 1080p to 4k.

 

[Poneros]

When you have roughly comparable output from both DLSS is providing around 2 to 3 times the performance uplift that FSR does.

How do you figure that? Need to hear what you're thinking of because honestly that sounds absurd bordering on the ridiculous. That would be what - DLSS performance vs FSR UQ? Sounds a tad bit optimistic for DLSS, since we know the costs. So you must like that DLSS image a lot more for that to make sense.

 

[Rroff]

How do you figure that? Need to hear what you're thinking of because honestly that sounds absurd bordering on the ridiculous. That would be what - DLSS performance vs FSR UQ? Sounds a tad bit optimistic for DLSS, since we know the costs. So you must like that DLSS image a lot more for that to make sense.

You need to look at more than the overhead from the process itself - hypothetically if one needs to render at a higher input resolution to match the other at a given output quality that is going to impact the final rendering performance, etc. amongst other considerations which will impact on the actual uplift vs native.

There is a lot of absurdity bordering on ridiculous in this thread but it certainly isn't coming from me - though no doubt it will be weeks and months down the line before that becomes more apparent as usual by which time people will move on to another topic.

 

[Hotwired]

He has a point though, what are you comparing to get 2-3x uplift.

 

[nvidiamd]

How do you figure that? Need to hear what you're thinking of because honestly that sounds absurd bordering on the ridiculous. That would be what - DLSS performance vs FSR UQ? Sounds a tad bit optimistic for DLSS, since we know the costs. So you must like that DLSS image a lot more for that to make sense.

Spoiler

Yeah that sounded pretty bad even if we compare the DLSS that ran on Nvidia shaders DLSS won't get twice the performance. Usually you get whatever the initial res FPS is +/- several frames.
So if you upscale from 1440p then you get native 1440p perf, if you upscale from 1080p you get native 1080p perf and so on.

 

[nvidiamd]

You need to look at more than the overhead from the process itself - hypothetically if one needs to render at a higher input resolution to match the other at a given output quality that is going to impact the final rendering performance, etc. amongst other considerations which will impact on the actual uplift vs native.

There is a lot of absurdity bordering on ridiculous in this thread but it certainly isn't coming from me - though no doubt it will be weeks and months down the line before that becomes more apparent as usual by which time people will move on to another topic.

So you say that because DLSS is looking better than FSR, you can use DLSS performance vs FSR ultra quality because they will give the same quality output. Which seems fair if the things are this way but the quality is a subjective opinion to a certain degree. You will never convince someone that his baby is uglier than your baby, unless there is a huge difference between them.

 

[nvidiamd]

FSR is also eco friendly so we will have a chance to save the planet:
https://twitter.com/CapFrameX/status/1408095491280211969

DLSS probably does the same thing though.

 

[Rroff]

He has a point though, what are you comparing to get 2-3x uplift.

For instance https://www.techpowerup.com/review/...per-resolution-quality-performance-benchmark/ at 1440p FSR UQ is giving around 28% average uplift over native for a ~3070 class card - look at benchmarks for DLSS Quality in CP2077, NMS or Watchdogs Legion, etc. it is around a 76-80% uplift over native and that is assuming that FSR UQ was as good as DLSS Quality mode which overall it isn't though some people might find the DLSS artefacts subjectively detract more from the experience.

Even stuff like the 3080 in CP2077 gets around 40-50% performance uplift at 1440p with DLSS Quality mode.

 

[humbug]

So you say that because DLSS is looking better than FSR, you can use DLSS performance vs FSR ultra quality because they will give the same quality output. Which seems fair if the things are this way but the quality is a subjective opinion to a certain degree. You will never convince someone that his baby is uglier than your baby, unless there is a huge difference between them.

When you're comparing visual quality that's near identical you're clutching at straws, DLSS 2 is better, but the difference is very minor.

 

[Rroff]

So you say that because DLSS is looking better than FSR, you can use DLSS performance vs FSR ultra quality because they will give the same quality output. Which seems fair if the things are this way but the quality is a subjective opinion to a certain degree. You will never convince someone that his baby is uglier than your baby, unless there is a huge difference between them.

I've been comparing best effort mostly - it would need a more in-depth analysis to work out accurately where each DLSS mode stacked up against an FSR mode.

 

[Nexus18]

We are talking about the costs in time/hours it takes implementing FSR or DLSS from a blank slate. Stating DLSS is just a switch is disingenuous when you know it took time and effort to implement in the base game engine in the first place. The concensus among developers and the tech press is that implementing FSR from scratch takes less time and work than implementing DLSS.

FSR is open source, available to a much wider user base, cross platform, easier (and cheaper) to implement from scratch. So I will predict it will overtake DLSS in number of games supported over the coming years. Probably much sooner.

EDIT: DLSS will of course till exists but FSR is here and here to stay.

Why though? Is there something stopping developers from using the plugin/easy way of implementing dlss??? If the hard part has been done for them then why would they go and try to do it from scratch when they can just do the easy way???

Not disputing that FSR is easy from the get go but to completely blank out that dlss is also just as easy, if not easier to enable/add in unreal engine, unity based games is silly when that is not the case as per developers comments.

And again, don't disagree that FSR will eventually take the lead in the amount of games it is in but as it is right now, dlss is in more games and in games which need all the perf. they can get.

 

[nvidiamd]

For instance https://www.techpowerup.com/review/...per-resolution-quality-performance-benchmark/ at 1440p FSR UQ is giving around 28% average uplift over native for a ~3070 class card - look at benchmarks for DLSS Quality in CP2077, NMS or Watchdogs Legion, etc. it is around a 76-80% uplift over native and that is assuming that FSR UQ was as good as DLSS Quality mode which overall it isn't though some people might find the DLSS artefacts subjectively detract more from the experience.

Even stuff like the 3080 in CP2077 gets around 40-50% performance uplift at 1440p with DLSS Quality mode.

You don't understand that it depends on each game, some games will have very low fps gains since they are not hard to run on the card anyway even at native. While others are very hard to run especially RT heavy games.

For example with the 3070 in CP 2077 at 1080p ultra no RT, you get 83 FPS. Which is pretty much what you get at 1440p ultra DLSS quality no rt.
This is how it works on every game, you get pretty much the performance you had on the initial resolution.

 

[Rroff]

You don't understand that it depends on each game, some games will have very low fps gains since they are not hard to run on the card anyway even at native. While others are very hard to run especially RT heavy games.

For example with the 3070 in CP 2077 at 1080p ultra no RT, you get 83 FPS. Which is pretty much what you get at 1440p ultra DLSS quality no rt.
This is how it works on every game, you get pretty much the performance you had on the initial resolution.

There are a huge range of scenarios based on resolution, GPU and game, etc. but taking an average the performance uplift from DLSS quality mode is substantially higher than FSR.