If DirectML could promise these kinds of performance gains this promises, I'd rather XSX games run at 1440p reconstructed to 4K, leaving more processing power for more intensive graphics. Running native seems like like a waste of processing power.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
Digital Foundry || Control vs DLSS: Can 540p Match 1080p Image Quality? Full Ray Tracing On RTX 2060?
- Thread starter ILikeFeet
- Start date
If DirectML could promise these kinds of performance gains this promises, I'd rather XSX games run at 1440p reconstructed to 4K, leaving more processing power for more intensive graphics. Running native seems like like a waste of processing power.
OP
OP
My point is that zoom isn't necessary for everybody to see the faults. I can see them with no zoom at all on my PC. But you're correct that for different people, and different setups, it may be an acceptable tradeoff. For example, if I were playing on my TV I probably would have a hard time noticing, due to increased viewing distance only partially offset by the bigger screen.
Unless Nintendo make the "Switch 2" screen physically a lot bigger, such high resolution is somewhat overkill. DLSS may be better than typical upscaling at preserving single-pixel detail, but on a 4K tablet screen, single-pixel detail is going to be tough to see anyway.
I think the better scenario would be to just upscale from 1080p, or 1440p, and use any extra resources to add higher polys, better materials, more particles, cool liquid simulations, and especially excellent and fully dynamic light and shadow. I think those sorts of things would enhance Nintendo's chosen graphical styles more than just adding lots of teeny details.
AMD's answer for machine-learning type work appears to be different. Instead of running such work on dedicated hardware, it runs on the normal general-purpose hardware, but at increased speed versus other types of work. At least, that's how Series X does it, and PS5 is the same chip family so liable to also. This is better than nothing, but worse than DLSS because it's even more of a tradeoff with other graphics work. (It might still make sense, though, since you save so much overhead dropping to one-quarter the resolution.)
I think it's more that AI upscaling just isn't a magic bullet. Some TVs have had AI upscaling for years (not realtime, done with a trained DSP). It's pretty hard to see the supposed advantages. Game engines do have more potential upside, though, since it's not totally post-process.
Thanks for the detailed reply <3
Yeah I play on a 50" TV but I sit maybe 4 feet from my display so it would be very difficult to tell.
I was more meaning for docked play on the Switch Pro/2. 4k would definitely be overkill on a 6" screen haha.
Interesting about AMD. Will have to see what they come up with. In truth 1440p or CB4k is great image quality for me and my set up on the PS4 Pro. I'm mostly excited about this for the next Switch as 720-900p really doesn't look great on a 4k screen even from 4 feet away!
This is amazing, not only for PC but also potentially the next Switch. There was always the question of how it would be able to handle next ports, this could be the answer.
Native resolutions are a fool's errand next generation (unless the game is basic and/or cross-gen).
Awesome tech. Cheers for the explanation.
Wonder if Switch (or its next model) could start using something like this.
How is that TOPs performance allocated on RTX cards? Could the 2060 do all those 200 Tops on the tensor cores alone without affecting general shader performance, whereas the entire XSX GPU could only manage half the TOPs?
No, NVIDIA drivers divert any FP16 calculations to the Tensor cores on RTX GPUs, no developer intervention is required. Also Far Cry 5 and Wolfesntein games use FP16 on any hardware that supports them, Vega, Turing, Volta, RDNA .. etc.
OP
OP
Seeing this stuff just increases my anticipation for the 30 series. If we get even a 40% jump over the 20 series, we could be looking at a host of performance options that give users even more options that PC gaming has ever offered. What an amazing time to observe all of this.
Someone dumb it down for me. If this is implemented in Switch 2 what exactly would it mean from a 3rd party support perspective?
Basically just one of the best forms of upscaling we've seen so far. You can render a game internally at half the final output resolution and it still looks shockingly good. On PC it's being touted as a way to use GPU intensive features like ray tracing without as big of a performance hit, and on a Switch 2 it could mean games looking much better in docked mode.
Games will look prettier with same old lower than low resolution.
Nvidia has "AI based" real time video upscaling on the latest Shield TV models, but it's not really the same as DLSS since it's not using a heap of training data of how the content is supposed to look at higher resolutions that it's trying to match.
The results are still pretty impressive by most accounts. Here's a video review.
You... do realize the Switch has a mode where it can hook up to a big ass TV, right?
With all the Switch successor talk it would be nice if someone can find out how many tensor cores are required for this and at what frequency. The RTX2060 has 240 but probaly can't expect a Switch successor to have that.
I was blown away by this when I went back to Control for the DLC.
It really feels like people are tuning this out, the blurry mess that was the first showcase of DLSS on Metro and Tomb Raider and whatever else is still in people's mind I guess.
I won't be surprised at all when this is the showcase feature of the RTX3000 launch and they conveniently ignore/footnote that you don't need the new gen for it.
It really feels like people are tuning this out, the blurry mess that was the first showcase of DLSS on Metro and Tomb Raider and whatever else is still in people's mind I guess.
I won't be surprised at all when this is the showcase feature of the RTX3000 launch and they conveniently ignore/footnote that you don't need the new gen for it.
OP
OP
games can be rendered at lower resolution and still have really good IQ. because it's not really the loss of texture resolution or visual effects that hurts Switch ports, but the low-ass resolution. this solves that problem
We will see. A switch successor's SoC will likely be built on a 7nm process, allowing for more cores, but not sure how many given that it will be a mobile chip instead of a dedicated GPU.
Laymen here... what's the significance of upscaling on dedicated tensor cores vs AMDs approach if nvidia's GPU is idle during the upscale anyway?
Is taking from the same pool of resources as all other rendering a big negative if the alternative is the GPU idling until the upscale?
I'm already planning to buy a 3000, but it would make me rush to a store and get one. Honestly.
On Switch 2, XSX and PS5 ports could run at 540p upscaled to 1080p undocked and at 1080p upscaled to 4K docked.
Wasn't something like this mooted in old Frostbite and CryEngine videos?
The difference is that the tensor cores are much more capable at the INT8/INT4 computations than the AMD CUs, which make the DLSS upscale phase faster. That seems to be the extent of the benefit.
Why hello there Switch 2. Say hi to 1080p undocked and 4K docked resolutions even potentially of PS5/XBX ports.
Also holy shit ... I actually like the look of the DLSS 540p (reconstructed to 1080p) better than the 1080p native.
That is a monstrous game changer for Switch 2, even at say 1.6 TF undocked and 3.2 TF docked ... 1.6 TF only having to handle a meager 540p resolution ... it's going to crunch that low of a pixel count like eating cake.
Also holy shit ... I actually like the look of the DLSS 540p (reconstructed to 1080p) better than the 1080p native.
That is a monstrous game changer for Switch 2, even at say 1.6 TF undocked and 3.2 TF docked ... 1.6 TF only having to handle a meager 540p resolution ... it's going to crunch that low of a pixel count like eating cake.
Last edited:
I'm wondering about that too. With DLSS the GPU should certainly be able to manage PS5 games. But it's not just the GPU. They'd need a lot of fast RAM, and an 8c/16t Zen 2 CPU. Maybe doable in 2022? But at what price...?
OP
OP
They don't need that, they just need a sufficiently fast CPU and ram for their targeted resolution. 8 Arm cores at a high enough clock and 8GB of LPDDR5 would be enough for ports.
I feel like I asked this question in a previous thread but didn't word it well: are amd working on similar upscaling tech/is this something the other consoles could benefit from at all?
I'll be honest, I prefer the look of DLSS 540p and 1080p (reconstructed to 1080p and 4K) over the 1080p and 4K native, lol.
It looks sharper and higher resolution to my eye and ultimately more eye pleasing to me. I mean yeah I guess you can become a tech nerd and scrutinize every pixel, but if this is the type of result they can pull off, the average gamer I doubt very much is going to be able to tell much of a difference or may even prefer the Nvidia DLSS-ed image.
I almost wonder if there would be much point to going even as high as 1080p docked for Switch 2 if they used this for really ambitious games. 900p would probably scale up to 1800p in that case and 1800p I think for most people is pretty damn close 4K. But rendering at an even lower native res means you can probably push visual fidelity itself even higher.
It looks sharper and higher resolution to my eye and ultimately more eye pleasing to me. I mean yeah I guess you can become a tech nerd and scrutinize every pixel, but if this is the type of result they can pull off, the average gamer I doubt very much is going to be able to tell much of a difference or may even prefer the Nvidia DLSS-ed image.
I almost wonder if there would be much point to going even as high as 1080p docked for Switch 2 if they used this for really ambitious games. 900p would probably scale up to 1800p in that case and 1800p I think for most people is pretty damn close 4K. But rendering at an even lower native res means you can probably push visual fidelity itself even higher.
OP
OP
DLSS is Nvidia's work. Microsoft has a similar ML upscaling solution and RDNA2 does support machine learning workloads, so who knows how that might do on Series X
My only issue with DLSS at the moment is how it is incredibly oversharpened.
I'm not sure if this just because gamers have a
terrible eye for that kind of thing and prefer it (so it's forced for extra BAMpop) or if it's an unfortunate side effect of the training
I'm not sure if this just because gamers have a
terrible eye for that kind of thing and prefer it (so it's forced for extra BAMpop) or if it's an unfortunate side effect of the training
Yeah I knew DLSS is a Nvidia solution but feels like these techs usually end up with wider adoption.
Will be interesting to see what this does this gen/whether these consoles not having this kind of tech proves to be a major weakness
I think the interesting point is that to start getting into "differences" you have to do 400-800 % zooms.
With normal 2160 checkerboarding, the overaveraging of pixel results makes the differences between a real native and checkerboarded result in side by sides noticably different in aggregate.
Also the fact that we are comparing and looking for the "which is better/ where does it fail" in side by sides with native and zooming in to tell the reasons why it is deficient while it also is 130% faster.
That is... rather unique.
It is 100% a togglable slider. They left it as is probably just because basic market research told them people like images sharp to the point of having ringing artefacts.
The sharpening halos are the problem as I see iot. The actual detail in inner surfaces in most shots which make it look different than native, is usually just because native has TAA overeveraging results in inner surfaces, turn off TAA in such shots and inner surface detail actually looks remarkably like DLSS.
One of the lead developers behind the technology said sharpness is actually scalable, they just didn't happen to include it in any games yet.