Then what on earth are you doing here?just kidding of course.
Indeed, although I am curious as to how things will progress moving forward, will Playstation iterate upon PSSR for the PS6 on the UDNA architecture, or will they pluck the latest iteration of FSR 4 or 5?
I do not agree. These new models, DLSS4 and FSR4, requiere at least 4 times more TOPS than what the Pro has. Once again, Cerny had to make it work under a certain budget and I think PSSR works well for what it was intended to.
Then what on earth are you doing here?
Indeed, although I am curious as to how things will progress moving forward, will Playstation iterate upon PSSR for the PS6 on the UDNA architecture, or will they pluck the latest iteration of FSR 4 or 5?
No, it is far more than that given that consoles can't afford to waste silicon and pass it onto the consumer while keeping costs down.
But it’s not even a Pro thread. This should just be about FSR4 (which is currently a PC-only thing). That the Pro exists in this discussion is the real mystery here (well, not if you read one of the first posts in here).
Do you equally object to DF comparing it to DLSS CNN and DLSS transformer model, then? As FSR4 isn't PC only, it is RDNA4 only.
no it doesn't struggle
I’m curious to know if you are aware of how the ML acceleration from RDNA4 compare to CDNA3? If they differ significantly do you expect UDNA/PS6 to follow which path?
CDNA supports larger matrix sizes and more precision formats. Also it has an additional register file to reduce VGPR pressure. I don't know what they are doing with matrix HW in gfx13 yet.
this is performance mode against quality mode, not actual native (so a 33% internal resolution reduction)
The general point was clearly lost on you, as a comparison to PSSR wasn’t the issue.
Eh, it's inevitable PSSR will come up in these AI upscaling discussions. It's actually a testament to PSSR's competence really.
I agree - comparisons are inevitable. My main point is how we get to these silly back-and-forth battles, with people needlessly trolling and shot-taking.
github.com
well games tested in the 2060 super benchmark and your examples don't have that much in common
I tested many games and I dont remember any game that would show just 10% relative difference between DLSSQuality and Performance. The RTX2060 has 10% relative difference almost in every game. DLSS transformer works on the RTX2060, but performance scaling is just not good enough. I would never use DLSS Performance if it would be only offering 10% performance boost over DLSS Quality.
what? it has 36% performance improvement in cyberpunk
Optiscaler already has an early build to mod FSR4 into DLSS games.
Release Nightly Builds · cdozdil/OptiScaler
Changes in the Last 24 Hours 7bff360 fixed kmt (cdozdil) a3135cd push kmt for testing (cdozdil) b14c1cd Added comments for FSR4 upgrade (cdozdil)
Optiscaler already has an early build to mod FSR4 into DLSS games.
Release Nightly Builds · cdozdil/OptiScaler
Changes in the Last 24 Hours 7bff360 fixed kmt (cdozdil) a3135cd push kmt for testing (cdozdil) b14c1cd Added comments for FSR4 upgrade (cdozdil)
The RTX2060 gained 8% to 27% going from DLSS Quality to DLSS Performance. This is based on 10 tests. I also did 10 tests and my card gained between 20% and 48%.
Performance looks so good that quality mode is almost redundant now. Not worth the hit, imo.The RTX2060 gained 8% to 27% going from DLSS Quality to DLSS Performance. This is based on 10 tests. I also did 10 tests and my card gained between 20% and 48%.
Nvidia offers support for DLSS4 and ray reconstruction on Turing GPUs, but the (performance) cost of using these features is obviously much higher compared to newer generations. It's still great that Nvidia has offered DLSS4 / RR support for the old Turing, but this technology obviously needs more tensor cores / AI resources in order to work at its best.
Alan Wake 2 - 21% relative difference
Cyberpunk - 27% relative difference
Final Fantasy Rebirth - 10% relative difference
GOW Ragnarok - 11% relative difference
Hogwarts Legacy - 13% relative difference
Horizon 2 - 8% relative difference
Indiana Jones - 16% relative difference
Stalker 2 - 16% relative difference
The First Descendant - 14% relative difference
The Last Of Us Remake - 19% relative difference
It's still worth running games at a higher resolution, because some effects like the resolution of screen space reflections, and not to mention RT are tied to internal resolution, and will look much worse when using lower DLSS modes. Also ray reconstruction is extremely sensitive to internal resolution. DLAA RR looks amazing (image look razor sharp and textures arnt filtered), but in DLSS Performance mode ray reconstruction makes the image look like a paint (filtered / overprocessed look).
I tested some of the games from that list, but I dont have all of them to test every single one.
Any comparisons out there of this?
One reason would be to take Nvidia's DLSS stick away and destigmatizing GPUs that aren't Nvidia - resulting in better competition and an more open inclusive standard for ML AI scaling on PC - lowering GPU prices.
It's only any good (as in acceptable trade off in image quality versus performance gain) in quality mode (4K not sure about any other output res)- anything below and it turns to absolute shit rapidly. DLSS3 suffered in quality in performance mode, but could work.in balanced mode and was pretty much on par with native in quality mode. The new upscalers can finally give acceptable results in performance mode.
I knew this was a PaintTinJr post before I scrolled up to see the name.
And that matters why - exactly ?
Thanks! Yeah RR still has a ways to go. Whichever genius figures out how to stop the boiling with upscaling should get a big fat raise. I think it’s just a matter of time.
