3 ms·

> Nvidia is already very keen on emphasizing DLSS as the solution to this problem, and AMD is likewise tauting its own tech as well. Personally I dislike DLSS

> Nvidia is already very keen on emphasizing DLSS as the solution to this problem, and AMD is likewise tauting its own tech as well. Personally I dislike DLSS so much: If I enable it in Cyberpunk 2077, there are some visible noise patterns that pop in and out of view especially near shadows on flat textures all the time, it looks horrible and distracting. And if I disable DLSS, if the game updates sometimes it enables it back so I see the horrible things again! Why do you need AI to upscale anyway if it looks this bad, why not just use Lanczos/bicubic/... upscaling? They are not only simpler and faster, but look better too, because they don't create the above mentioned horrible visible noise patterns. Then they have the inter-frame thing that causes input lag, another bad degradation. Then they 'solve' input lag by touting a vendor lock-in-sounding feature (Nvidia Reflex), while the real solution to input lag is not add vendor lock-in feature but ensure you compute 1 frame at the time and do so fast (which is of course not simple, but is a matter of optimization and not of brand-named features, and perhaps not generating and inserting additional frames would help). Then there is the fact that DLSS is something that needs to be supported by games themselves: why? This again makes it sounds vendor-lock-in-ish because they can have it "exclusively" in some games. But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me. The graphics card should have enough info to do that by having all the pixels. I mean, it was always possible to play games at different resolutions (DLSS is basically also like playing the game at lower resolution), and there always has been some upscaling required for that (whether it's the monitor or the graphics card doing it). This never required special engine support.

gjsman-1000 3y ago
> This again makes it sounds vendor-lock-in-ish because they can have it "exclusively" in some games. But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me Because DLSS is a AI-based solution with customized settings and inputs for each game. (The “Deep Learning SuperSampling” is right in the title.) This is also why it requires NVIDIA’s bespoke Tensor cores to function. DLSS 2.0 removed most per-game optimization but there’s still some support work required. AMD’s FSR is basically a traditional upscaling solution, just with low overhead. FSR runs fine on NVIDIA and even the Nintendo Switch - but almost all reviewers agree DLSS is superior in most games. Except, perhaps, Cyberpunk 2077. But that’s not a fair metric to measure DLSS by because everything about that game has been a technical disaster.
- Aardwolf 3y ago
  > This is also why it requires NVIDIA’s bespoke Tensor cores to function. Any idea why this needs bespoke cores? GPU's were already used for AI and crypto mining before this, without bespoke cores for 1 specific application > The “Deep Learning SuperSampling” is right in the title With that name, it could have been based off of a single pre-training on many games, to work on games in general
  - gjsman-1000 3y ago
    As far as I know, Tensor cores are a little like Raytracing cores. Precision, unlike for mining or science, is not valued and they are allowed to do sloppy math within a tolerance in exchange for much, much greater speed. > With that name, it could have been based off of a single pre-training on many games, to work on games in general Well, I’m not NVIDIA, so I don’t know if you could. I just know AMD tried and the results were almost universally considered inferior.
tomnipotent 3y ago
> They are not only simpler and faster, but look better too I disagree. It's not even close how much better DLSS looks than traditional upscaling. I've run into very few noticeable artifacts from DLSS that are distracting, and this seems to be the general consensus among most consumers. > DLSS is something that needs to be supported by games themselves: why? It requires motion vectors and depth buffers in addition to the low res aliased images.
- SirMaster 3y ago
  I wish I shared the same sentiment as most consumers. DLSS looks poor to me and I have yet to actually keep it enabled.
- Retric 3y ago
  The issue is generally looking better isn’t enough. The goal is immersion and even relatively rare artifacts destroy that feeling so DLSS is simply worse than playing at a lower resolution that I’m going to get used to in a few minutes. This is the same reason 60FPS in modern games isn’t “enough” while it felt buttery smooth on old console fighting games. Consistent FPS is vastly more important than average FPS.
  - tomnipotent 3y ago
    > The goal is immersion Here I was thinking the goal was entertainment. The reality is that we're dealing with trade-offs, and one of them is reducing resolution to improve FPS. If I can't play at native resolution with adequate FPS, the options are to reduce quality or resolution. I personally prefer to keep quality as high as possible, even if that means a lower resolution. If you reduce resolution, the first thing you'll notice is that the result looks blurry, or otherwise just less sharp thanks to whatever is doing the default upscaling. DLSS lets me reduce resolution while retaining a sharp image that looks like native resolution. That's a win-win if you don't enjoy 4k@30fps. > This is the same reason 60FPS in modern games isn’t “enough” There are hundreds of millions of console users that disagree. It's great to get a billion FPS, but the reality is that outside of a few people most gamers just don't notice the difference after a certain point.
enragedcacti 3y ago
> Then there is the fact that DLSS is something that needs to be supported by games themselves: why? DLSS requires motion vectors, depth buffer, and jitter in addition to the rendered frame. DLSS 2.0 is effectively Temporal Anti-aliasing (TAA) with deep learning replacing the heuristic approach of TAA implementations and thus needs as much integration as TAA does. https://old.reddit.com/r/hardware/comments/fvgf7d/how_dlss_20_works_for_gamers/ https://old.reddit.com/r/hardware/comments/fvgf7d/how_dlss_2...
rand_flip_bit 3y ago
> But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me. The graphics card should have enough info to do that by having all the pixels. Modern upscaling solutions (eg TAA) require not just the pixel outputs but various inter-frame/metadata such as depth, normals, motion vectors, last frame, etc. DLSS is more or less just the spiritual successor to TAA with machine learning sprinkled in. Integration with the game engine is needed because the non-color metadata needs to be provided to the DLSS and how each engine computes those things is a bit different.
charcircuit 3y ago
>while the real solution to input lag is not add vendor lock-in feature but ensure you compute 1 frame at the time and do so fast No, Nvidia Reflex is a step in the right direction. You want to sample input at the latest possible time and then create a frame for that input. Sampling input at the beginning of a frame is the worst time you can choose as that will guarantee at least a frame of latency. If you are able to render in 20% of the time of a frame you want to sample the input and then start rendering 80% of the way through the current frame that way you are only add 20% of a frame of latency at a minimum in this part of the system. This is how things are done in VR, but the location of input devices are also exterpolated to predict where they should be located at time they will be displayed.
- Aardwolf 3y ago
  > but the location of input devices are also exterpolated That's impossible, if you change your movement it will mispredict and thus cause more lag, while exactly when you change is when it should feel the best. Consistency feels better. Just like no magic gizmo can reduce audio lag other than have actual proper low latency standards (for bluetooth audio): it can never predict when you will hit the button to fire your weapon, etc..., in games
  - charcircuit 3y ago
    >That's impossible, if you change your movement it will mispredict and thus cause more lag It will not cause more lag, it will just not be accurate to where your cursor actually should be when the frame gets displayed. For people with strong computers you would only need to predict a few ms forward in time which should be possible to get really accurate. >it can never predict when you will hit the button to fire your weapon That is not what is happening. What is happening is that time between registering the click and the frame being displayed. The prediction I'm talking about would be predicting subframe mouse movements. This behaviour is more for smooth mouse look. For shooting you want to have a precise timestamp of when the click happened so that you can see if at that moment in time the shot would have hit. The position where you are looking when shooting and the position where you are looking on screen are different things.
I_Am_Nous 3y ago
>Then there is the fact that DLSS is something that needs to be supported by games themselves: why? Old news, but in a similar vein -- games that incorporated PhysX back in the day required nVidia hardware to make it work. I imagine its a similar issue for PhysX and DLSS, in that the game engine has to know to expect PhysX calculations and have something to run those calculations, and then use those calculations for debris or ragdoll physics. It could be turned on or off (so AMD cards could still play the game with a different physics backend). It sounds like another case of nVidia making a cool technology and game devs choosing to use that cool tech in their game (or not, looking at you Starfield).
sosodev 3y ago
You’re definitely in the minority. Most people agree that DLSS is the best upscaling tech available right now. It produces an image that is genuinely similar to a higher native resolution with fairly minor artifacts. It probably could run on other GPU but Nvidia isn’t the type to just give tech to their competitors.

gjsman-1000 3y ago

> This again makes it sounds vendor-lock-in-ish because they can have it "exclusively" in some games. But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me Because DLSS is a AI-based solution with customized settings and inputs for each game. (The “Deep Learning SuperSampling” is right in the title.) This is also why it requires NVIDIA’s bespoke Tensor cores to function. DLSS 2.0 removed most per-game optimization but there’s still some support work required. AMD’s FSR is basically a traditional upscaling solution, just with low overhead. FSR runs fine on NVIDIA and even the Nintendo Switch - but almost all reviewers agree DLSS is superior in most games. Except, perhaps, Cyberpunk 2077. But that’s not a fair metric to measure DLSS by because everything about that game has been a technical disaster.

Aardwolf 3y ago
> This is also why it requires NVIDIA’s bespoke Tensor cores to function. Any idea why this needs bespoke cores? GPU's were already used for AI and crypto mining before this, without bespoke cores for 1 specific application > The “Deep Learning SuperSampling” is right in the title With that name, it could have been based off of a single pre-training on many games, to work on games in general
- gjsman-1000 3y ago
  As far as I know, Tensor cores are a little like Raytracing cores. Precision, unlike for mining or science, is not valued and they are allowed to do sloppy math within a tolerance in exchange for much, much greater speed. > With that name, it could have been based off of a single pre-training on many games, to work on games in general Well, I’m not NVIDIA, so I don’t know if you could. I just know AMD tried and the results were almost universally considered inferior.

tomnipotent 3y ago

> They are not only simpler and faster, but look better too I disagree. It's not even close how much better DLSS looks than traditional upscaling. I've run into very few noticeable artifacts from DLSS that are distracting, and this seems to be the general consensus among most consumers. > DLSS is something that needs to be supported by games themselves: why? It requires motion vectors and depth buffers in addition to the low res aliased images.

SirMaster 3y ago
I wish I shared the same sentiment as most consumers. DLSS looks poor to me and I have yet to actually keep it enabled.
Retric 3y ago
The issue is generally looking better isn’t enough. The goal is immersion and even relatively rare artifacts destroy that feeling so DLSS is simply worse than playing at a lower resolution that I’m going to get used to in a few minutes. This is the same reason 60FPS in modern games isn’t “enough” while it felt buttery smooth on old console fighting games. Consistent FPS is vastly more important than average FPS.
- tomnipotent 3y ago
  > The goal is immersion Here I was thinking the goal was entertainment. The reality is that we're dealing with trade-offs, and one of them is reducing resolution to improve FPS. If I can't play at native resolution with adequate FPS, the options are to reduce quality or resolution. I personally prefer to keep quality as high as possible, even if that means a lower resolution. If you reduce resolution, the first thing you'll notice is that the result looks blurry, or otherwise just less sharp thanks to whatever is doing the default upscaling. DLSS lets me reduce resolution while retaining a sharp image that looks like native resolution. That's a win-win if you don't enjoy 4k@30fps. > This is the same reason 60FPS in modern games isn’t “enough” There are hundreds of millions of console users that disagree. It's great to get a billion FPS, but the reality is that outside of a few people most gamers just don't notice the difference after a certain point.

enragedcacti 3y ago

> Then there is the fact that DLSS is something that needs to be supported by games themselves: why? DLSS requires motion vectors, depth buffer, and jitter in addition to the rendered frame. DLSS 2.0 is effectively Temporal Anti-aliasing (TAA) with deep learning replacing the heuristic approach of TAA implementations and thus needs as much integration as TAA does. https://old.reddit.com/r/hardware/comments/fvgf7d/how_dlss_20_works_for_gamers/ https://old.reddit.com/r/hardware/comments/fvgf7d/how_dlss_2...

rand_flip_bit 3y ago

> But nothing about upscaling, nor frame generation, sounds like something that requires game engine support to me. The graphics card should have enough info to do that by having all the pixels. Modern upscaling solutions (eg TAA) require not just the pixel outputs but various inter-frame/metadata such as depth, normals, motion vectors, last frame, etc. DLSS is more or less just the spiritual successor to TAA with machine learning sprinkled in. Integration with the game engine is needed because the non-color metadata needs to be provided to the DLSS and how each engine computes those things is a bit different.

charcircuit 3y ago

>while the real solution to input lag is not add vendor lock-in feature but ensure you compute 1 frame at the time and do so fast No, Nvidia Reflex is a step in the right direction. You want to sample input at the latest possible time and then create a frame for that input. Sampling input at the beginning of a frame is the worst time you can choose as that will guarantee at least a frame of latency. If you are able to render in 20% of the time of a frame you want to sample the input and then start rendering 80% of the way through the current frame that way you are only add 20% of a frame of latency at a minimum in this part of the system. This is how things are done in VR, but the location of input devices are also exterpolated to predict where they should be located at time they will be displayed.

Aardwolf 3y ago
> but the location of input devices are also exterpolated That's impossible, if you change your movement it will mispredict and thus cause more lag, while exactly when you change is when it should feel the best. Consistency feels better. Just like no magic gizmo can reduce audio lag other than have actual proper low latency standards (for bluetooth audio): it can never predict when you will hit the button to fire your weapon, etc..., in games
- charcircuit 3y ago
  >That's impossible, if you change your movement it will mispredict and thus cause more lag It will not cause more lag, it will just not be accurate to where your cursor actually should be when the frame gets displayed. For people with strong computers you would only need to predict a few ms forward in time which should be possible to get really accurate. >it can never predict when you will hit the button to fire your weapon That is not what is happening. What is happening is that time between registering the click and the frame being displayed. The prediction I'm talking about would be predicting subframe mouse movements. This behaviour is more for smooth mouse look. For shooting you want to have a precise timestamp of when the click happened so that you can see if at that moment in time the shot would have hit. The position where you are looking when shooting and the position where you are looking on screen are different things.

I_Am_Nous 3y ago

>Then there is the fact that DLSS is something that needs to be supported by games themselves: why? Old news, but in a similar vein -- games that incorporated PhysX back in the day required nVidia hardware to make it work. I imagine its a similar issue for PhysX and DLSS, in that the game engine has to know to expect PhysX calculations and have something to run those calculations, and then use those calculations for debris or ragdoll physics. It could be turned on or off (so AMD cards could still play the game with a different physics backend). It sounds like another case of nVidia making a cool technology and game devs choosing to use that cool tech in their game (or not, looking at you Starfield).

sosodev 3y ago

You’re definitely in the minority. Most people agree that DLSS is the best upscaling tech available right now. It produces an image that is genuinely similar to a higher native resolution with fairly minor artifacts. It probably could run on other GPU but Nvidia isn’t the type to just give tech to their competitors.