I always find that "better than native" hilarious, but of course since what looks good is psychological, a matter of taste, it's not even automatically incorrect. Cameras (smartphone camera software) can have all kinds of filters to supposedly make an image better. Some simple things in Photoshop can potentially make images look better in human eyes. However, the fundamental problem is that game video is artificially created or reproduced (for example prerendered or recorded cutscene footage) in the first place. So, if a graphics card driver can make it look better, why didn't the game itself already make it look like that? And that's exactly where lies the problem in my opinion: If the driver makes the game look significantly/fundamentally different, is it not different from what the game studio intended? Of course a gamer's random screen might benefit from individual adjustment, which is not something game devs could easily handle, but automatic settings in a graphics card driver are no wiser, either.
Usually by "better than native" they mean better than the crappy TAA implemented in most games these days.
I always find that "better than native" hilarious, but of course since what looks good is psychological, a matter of taste, it's not even automatically incorrect.
Think lower native resolutions with AA and sharpening added, that is part of what FSR 2 includes with it's temporal approach, so it can indeed look better than native.
According to the claims it seems like AMD got their algorithms done right this time.
Might still require game developers to do some part of the work too, to get it working properly with given game engines.
Think lower native resolutions with AA and sharpening added, that is part of what FSR 2 includes with it's temporal approach, so it can indeed look better than native.
According to the claims it seems like AMD got their algorithms done right this time.
Might still require game developers to do some part of the work too, to get it working properly with given game engines.
Require game devs to do some part of the work? game devs can't even make their game work right without stuttering these days 😀
What is better.. AI to upscale from a low res taking up 25% silicon space with tensor cores.
Or a extra 25% in shader compute units for a increase in rasterization, Cuda cores.
I wonder..
Has anyone ever done the math ?
What is better.. AI to upscale from a low res taking up 25% silicon space with tensor cores.
ai that reconstructs on tensor cores that take up 3% silicon
lol,25%
Martin5000:
Has anyone ever done the math ?
have you done the math ?
https://twitter.com/locuza_/status/1057510611129679873?lang=en
it was under 15% for both rt cores+tensor cores on turing
ampere has half the tensor core count per sm compared to turing, 4 vs 8 per sm.
so take half of half of 15%,and see what it does in 2.3.x version to performance increase for image quality change.
I'm quite sure the fp32 to rt/tensor ratio is even lower than that ~3% since ampere's int32 units can do fp32 now unlike turing.
another fsr thread,another misinformation festival.what's new.
ai that reconstructs on tensor cores that take up 3% silicon
lol,25%
I had no idea that the tensor cores were so small.
3% of the chip for 30% extra performance is amazingly good.
cucaulay malkin:
it was under 15% for both rt cores+tensor cores on turing
ampere has half the tensor core count per sm compared to turing, 4 vs 8 per sm.
so take half of half of 15%,and see what it does in 2.3.x version to performance increase for image quality change.
I'm quite sure the fp32 to rt/tensor ratio is even lower than that ~3% since ampere's int32 units can do fp32 now unlike turing.
I thought the RT+Tensors occupied more space. Closer to the 25% mark.
13% of chip space for RT is much lower than what I expected.
But I would gladly give up those RT units for getting 12% more shaders, cache, etc.
I just don't care much about RT, because I almost never use it. But getting some 10% extra performance on all games would be a good trade off.
I had no idea that the tensor cores were so small.
3% of the chip for 30% extra performance is amazingly good.
I thought the RT+Tensors occupied more space. Closer to the 25% mark.
13% of chip space for RT is much lower than what I expected.
But I would gladly give up those RT units for getting 12% more shaders, cache, etc.
I just don't care much about RT, because I almost never use it. But getting some 10% extra performance on all games would be a good trade off.
In Dying Light 2 the rtx2070s/3060 with rt ultra and dlss quality at 1080p is averaging 30fps. No wonder you are never using becouse if you dont have 3080 type of card its not capable really.
In Dying Light 2 the rtx2070s/3060 with rt ultra and dlss quality at 1080p is averaging 30fps. No wonder you are never using becouse if you dont have 3080 type of card its not capable really.
Exactly, it's pointless to put RT units on cards like these. Or lower end.
Set fog to medium (often indistinguishable), don't use RT reflections and useless RT flashlight and you're getting the fps I mentioned (in starting area less, in city area more). The game is perfectly playable performance-wise with 3060 RT GI AO shadows + DLSS if you don't burn performance for dumb settings. Not exactly news that cranking up everything to max isn't the best idea with lower medium tier card (or anything else).
I've used RT in lots of titles on the 3060 in 1440p and found it enjoyable enough to keep it enabled. With gimped cards like GTX Turing there'll never be progress, even RDNA2 AMD APUs have RT...
ai that reconstructs on tensor cores that take up 3% silicon
lol,25%
have you done the math ?
https://twitter.com/locuza_/status/1057510611129679873?lang=en
it was under 15% for both rt cores+tensor cores on turing
ampere has half the tensor core count per sm compared to turing, 4 vs 8 per sm.
so take half of half of 15%,and see what it does in 2.3.x version to performance increase for image quality change.
I'm quite sure the fp32 to rt/tensor ratio is even lower than that ~3% since ampere's int32 units can do fp32 now unlike turing.
another fsr thread,another misinformation festival.what's new.
pls identify it pls be interesting to know.
cheers thanks.
Until we get better smaller pixels then the Coastline Paradox will always be a factor.
So jaggies will always be there but will continue getting smaller and smaller as pixel density increases.
I think future monitors will have much better PPI and this will help along with micro pixels.
Except I said myself that DLSS is bugged in DL2 in another thread. This may be shocking to you, but there are people who don't have judgement clouded based on their installed hardware. But don't bother if accepting this would be too traumatic.
With dlss perfromance at 1440p maybe check the link @Denial posted above. But yeah, setting fog to low gain around 10fps but still performance is bad.
Personally if the minimum fps drop below 60 fps too often i consider this not playable or at least not enjoyable. Can't play games anymore if i don't have a stable 60+ fps which obviously wont happen if the average is around 60.
Probably this part on the right. I have no idea where 3% is coming from though. Seems like a made up number.