• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

How upscaling tech compares, and which ones will prevail.

Sega Orphan

Banned
With the trends in graphics moving away from raw horse power to smarter more efficient ways of producing graphics fidelity, upscaling has become one of the most important parts of next gen graphics rendering.

So far we have the following.
Checkerboarding.
AMDs FSR
Nvidias DLSS
Unreal Engines Temporal Super Resolution
And in the future we have Intel's XeSS and Microsoft is going to be bringing out its own upscaling tech using Direct ML on the XSX/S.

Without a doubt DLSS has shown what can be accomplished using Machine Learning, as has TSR in non DLSS titles.
As well as that, the inclusion of checkerboarding on the PS4 Pro also allowed it to perform better than its GPU would alone.
We have even seen some games use two methods of upscaling, as I think Returnal did.

The issue is that some of the tech like DLSS requires unique hardware and is locked to one GPU maker. While ML is going to continue with Nvidia, and MS will do something with the Xbox Series consoles, as will Intel with XeSS, the software solutions such as FSR and TSR are improving all the time, and may get to a point where it's as effective as DLSS and ML, but without the need for hardware and training by super computers.

What other tech is out there, and where do you think the future lies for upscaling?
Which method gives the better results?
 

Tarkus98

Member
I think Insomniac uses something called temporal injection? Seemed to work well in R&C and Spider-Man games. Or was that for ray tracing. LOL.
 
Last edited:
I think Insomniac uses something called temporal injection? Seemed to work well in R&C and Spider-Man games. Or was that for ray tracing. LOL.
It’s their temporal upscaling technique and it’s the best in the business. Best upscaling technology that doesn’t use machine learning.

They’ve been working on reconstruction techniques a LONG time, since ratchet on ps3 which was sub hd reconstructed to hd with msaa samples.

Too bad their games are sour these days…
 
Last edited:
4ITv7Se.gif
 

Hugare

Member
DLSS is the best, but it's not hardware agnostic, so it wont be the leader in terms of adoption

I can see the UE 5 sollution being the most popular. It's damn good. When you turn it on/off during the Matrix demo, you can see how great it is.
 

S0ULZB0URNE

Member
I like whatever Sony's teams are using and expect it to evolve greatly when the nipple is taken out of the PS4's mouth.
 

kikkis

Member
I am but sceptical about superiority of dlss once fsr 2.0 hits town. I just don't see ml doing miracles in just the small millisecond or so it has time to do its thing. I think the temporal component of fsr 2 will bring it up to bar.
 
I am but sceptical about superiority of dlss once fsr 2.0 hits town. I just don't see ml doing miracles in just the small millisecond or so it has time to do its thing. I think the temporal component of fsr 2 will bring it up to bar.
Just like FSR 1.0, I'll wait on the results. (Same with DLSS 1.0, but it's punched way above it's weight over time)
 

Sega Orphan

Banned
I like whatever Sony's teams are using and expect it to evolve greatly when the nipple is taken out of the PS4's mouth.
Sony is kinda weird. They had their ICE teams brought together during the PS3 days to develop common technologies for Sony's internal studios to use, including game engines. However all of Sony's top studios (GG, SMS, Sucker Punch, ND and Insomniac) all use their own internal game engines, and they also develop their own tech like Insomniac has with their Temporal Injection.
 

Sega Orphan

Banned
Do we have an idea about how taxing the software upscaling is on the GPU? The various types of anti aliasing techniques had different costs to use. I am assuming the reconstruction techniques also have a cost associated.
I obviously less costly than to do it native, but maybe one is far more efficient than the others.
I know with the XSX and it's Machine Learning using it gives a 3 - 10 times performance improvement over stock.
 
I know with the XSX and it's Machine Learning using it gives a 3 - 10 times performance improvement over stock.
Maybe I'm misunderstanding your phrasing here. 3 to 10x performance improvement over stock. So a game running at say 40fps could technically hit 120 or higher using ML upscaling? Is that what you mean? DLSS 2.0 is typically a 30+ % improvement.
 

Sega Orphan

Banned
Maybe I'm misunderstanding your phrasing here. 3 to 10x performance improvement over stock. So a game running at say 40fps could technically hit 120 or higher using ML upscaling? Is that what you mean? DLSS 2.0 is typically a 30+ % improvement.
It's from a Microsoft presentation at Hot Chips I think.
gIy3rQN.jpg
 

Sega Orphan

Banned
Ya that doesn't really help :messenger_winking:

Sounds more like a 3-10x performance improvement on their ML by having it HW accelerated. The actual impact on framerate ... impossible to tell from that statement.
I'm thinking the same.
By running certain calculations on the lower precision (Int8, Int4 and FP16) for their ML it saves 3 X 10 times by having to use FP32.
We would have to see what the real world results are when, or if, it is released.
 

Sega Orphan

Banned
If FSR 2.0 works as shown today it will prevail due to it working on everything.
Yeah, it's a massive improvement over 1.0
If they can continue to improve it further then it will be one of the most used solutions.
With the potential of FSR 2.0, it makes me think that AMD might be bypassing any sort of tensor cores in future RDNA cards.
 
Last edited:
Top Bottom