Bondrewd
Veteran
Intel ARC GPUs, Xe Architecture for dGPUs [2018-2022]
- Thread starter DavidGraham
- Start date
-
- Tags
- intel
- Status
FSR? FSR didnt start anything. Games supported upscaling since years. FSR is a marketing tool which will die shortly.
An open source AI driven approach will have the same problems like DLSS. And it has to perform at least as good on nVidia hardware as DLSS.
Far more likely to be supplanted by FSR 2.0 next year which will be TSS with some AI as well possibly.
There won't be an "open source IHV agnostic" approach for upscaling since such upscaling will always be made by some vendor and will run differently on different h/w.
Bondrewd
Veteran
Nope.
1.0 will live on etc etc.
I neither know what mining algorithms do, nor did i follow NVs tricks to throttle mining perf., but: Is it possible to throttle a certain algorithm, while not affecting any other algorithms, on a general purpose processor? I'd naively guess the answer is no.
No fair comparison because they did upscale by a factor of 4, while current games methods use at most 2. So the fb method generates at least 4 times more pixels.
Also 4x upscale seems really much harder than just 2. I was very impressed from the results, but idk how much temporal subpixel information they use.
Last edited:
RDNA3+ will most likely follow that route, to implement AI hardware acceleration, not just for reconstruction tech.
Sure, it will "live on", it's on Github after all.
Doesn't mean that many games will be using it instead of relying on alternatives which will provide better IQ at the same performance.
Bondrewd
Veteran
Nope.
Matrix engines are forevah chained to CDNA CU/WGP variants.
Of course they will.
It's dumb simple which is the ~point~.
You always trade perf for IQ.
Haven't seen reports of perf regression from LHR GeForces. As long as it's a very specific algo and it's enough to throttle it during a few minutes, it should work. It's not like you'd have to detect a power virus instantly in order to save the card.
FXAA is dumb simple. Could you please name a game from 2021 which is using it?
Bondrewd
Veteran
Of course.
How is this even relevant to upscaling solutions.
Why are you moving the goalies; dats not fun.
I trust you'll make the connection, eventually.
Bondrewd
Veteran
Why even.
Yes FSR 1.0 is the antichrist and shall be banished from this world or w/ever helps you sleep at night.
Sounds i could calculate one frame of PacMan every second to crack mining detection >
Have not seen this slide, so thanks, but I am not arguing with the "conceptual illustration purposes only" slides.
I am talking about the issues with low precision modes, something which nobody would want to highlight in marketing slides or in the academia papers on "low precision" research.
As for perf, DP4A takes 2.2x more time, this translates into something like 3-4x more time on NN's execution, matches quite well with expectations, i.e. theoretical difference in ops numbers between DP4A and XMX.
These bars are useless though, there is 0 information on actual frame times, so whether that DP4A mode is productizable for real games has to be proven, a lot of issues to solve with precision, performance, post-processing and generalization.
Will be entertaining to see how open sourcing NN's training environment with data sets and all related quirks will accelerate adoption, there are so many data scientists with graphics engineering skills in game development that everybody would love to learn all that stuff immediately.
DLSS performance scales with output resolution, not input, 720p upscaling to 4K takes the same time as 1440p upscaling to 4K. The question is whether FB method is any different in this regard.
It's not.
TSR handles arbitrary input resolutions for example. It's just that quality at 25% rendering resolution is beyond bad.
FB method wasn't good with such input resolution either, it produced tons of wobbling on edges, something nobody would ever want to see in VR since any temporal instability causes motion sickness in VR.
D
Deleted member 13524
Guest
I think the comparison of FSR to FXAA as if it was a "bad thing" and how "it will quickly be forgotten" is really funny. Only one year after having its source code released to the public, FXAA was the most widely used form of AA in the whole market, due to how resource-efficient it was on the 7th-gen console GPUs and games. And at the time is was considered a very important development to achieve better looking games.
Even funnier is the fact that FXAA was developed by Timothy Lothes whom, according to AMD, was also the "main implementer of FSR" while working there.
So yeah, let's all hope that an IHV-agnostic upscaling solution like FSR "fails" as much as FXAA did 10 years ago.
Even funnier is the fact that FXAA was developed by Timothy Lothes whom, according to AMD, was also the "main implementer of FSR" while working there.
So yeah, let's all hope that an IHV-agnostic upscaling solution like FSR "fails" as much as FXAA did 10 years ago.
I'm pretty sure they could find a way to neuter SHA instruction throughput in the hardware. Even if some games may use it (GPU-based data decompression / decryption? is that ever used in the PC?), I doubt it would cause any meaningful performance impact.
And the LHR are only using a firmware+driver hack. It's just an afterthought of sorts.
My guess is FSR 1.0 will eventually be enabled through drivers, with FSR 2.x/3.x using linear + non-linear neural networks and probably needing more developer intervention.
- Status
Similar threads
