TrebleShot
Member
In English please for simpletons like me.
DLSS / AI Scaling and next gen consoles
- Thread starter THE DUCK
- Start date
DLSS is an overhyped TAA solution that existed for 10 years. It's just an evolution of existing software tech.
Jaguar Victory
Member
DirectML – Xbox Series X supports Machine Learning for games with DirectML, a component of DirectX. DirectML leverages unprecedented hardware performance in a console, benefiting from over 24 TFLOPS of 16-bit float performance and over 97 TOPS (trillion operations per second) of 4-bit integer performance on Xbox Series X. Machine Learning can improve a wide range of areas, such as making NPCs much smarter, providing vastly more lifelike animation, and greatly improving visual quality.
Defining the Next Generation: An Xbox Series X|S Technology Glossary - Xbox Wire
[Editor’s Note: Updated on 10/21 at 11AM to ensure it is now reflective of the capabilities across both of our next-gen Xbox consoles following the unveil of Xbox Series S.] As we enter a new generation of console gaming with Xbox Series X and Xbox Series S, we’ve made a number of technology...
news.xbox.com
Last edited:
IShowUBasics
Member
exactly. Noone needs that dlss meme. the fact that only control is shown every time someone talks about dlss just shows how bad and niche it actually is. In screenshots it obviously looks good but once its in motions its so bad looking.
regawdless
Banned
Can you elaborate what is bad looking about it in motion? Currently playing - of course - Control in 4k DLSS, and it looks very good. At least to me.
JimboJones
Member
I remember Microsoft saying they could use it with textures, basically they could have textures stored at quarter of the resolution (or something to that effect) and upscale them with AI and it looked just as good and it saved in bandwidth.
So not DLSS but an interesting use, article says they where experimenting with it so who knows if it will ever end up in use.
https://www.google.com/amp/s/wccfte...-textures-to-be-ai-upscaled-in-real-time/amp/
So not DLSS but an interesting use, article says they where experimenting with it so who knows if it will ever end up in use.
https://www.google.com/amp/s/wccfte...-textures-to-be-ai-upscaled-in-real-time/amp/
Last edited:
reezoo
Member
How good is dlss for doubling resolution say jumping from 1080p to 4K?
Checkerboard rendering is god send in this scenario as it requires close to double processing power instead of quadruple, reason why PS4 Pro was feasible.
There is always a possibility dlss might not give good result in some cases as it is black box reconstruction technique. While checkerboard gives predictable results.
Checkerboard rendering is god send in this scenario as it requires close to double processing power instead of quadruple, reason why PS4 Pro was feasible.
There is always a possibility dlss might not give good result in some cases as it is black box reconstruction technique. While checkerboard gives predictable results.
Last edited:
M1chl
Currently Gif and Meme Champion
What? No, they are training model with 16k image of the games which support this tech and it reconstruct the picture from that trained model. At least in DLSS2.0...
JimboJones
Member
Not that it's bad but I think DLSS 2.0 generally give much better results than the checkerboard rendering used on PS4
https://www.google.com/amp/s/www.eu...image-reconstruction-death-stranding-face-off
Phrixotrichus
Banned
Utter nonsense.....
Show me any TAA in any game that even gets within a mile of what DLSS does.
DLSS can reconstruct a 2160p image from a 1080p source and look indistinguishable from native resolution, or even better with small details.
No TAA solution in any game, ever, managed to create more than a muddy mess in that situation.
And while it is software...are you seriously denying that software with dedicated silicon runs several orders of magnitudes faster?!?
You`re announcing that no car will ever be able to do the quarter mile in under a minute while standing in front of a Ferrari. You do realize that DLSS 2.0 is already out there together with a giant number of comparison videos and analytics which all invalidate your claims?
Last edited:
Thugnificient
Banned
That's false. Control makes the best usage of it and there are like 5 games with DLSS 2.0.
Senua
Member
Overhyped my gonads. It looks basically identical to native, sometimes better because TAA is a blur fest. and best of all it saves a shit ton on performance.
Win win.
Last edited:
Thugnificient
Banned
The big problem is the dishonesty coming from either side and the utter lack of objectivity regarding the technology.
On one hand, you have people dismissing a technology with proven results (albeit on a limited number of games) with ridiculously false claims because it is not coming from their favorite manufacturer.
On the other hand you have people stating how much of a gamer changer it is despite the tech being 4 years old and looking impressive only in a handful of titles.
Almost none of the posts here is coming from a place seeking to have an informed discussion. It's mostly fanboy propaganda that undermines an otherwise potentially very interesting conversation.
On one hand, you have people dismissing a technology with proven results (albeit on a limited number of games) with ridiculously false claims because it is not coming from their favorite manufacturer.
On the other hand you have people stating how much of a gamer changer it is despite the tech being 4 years old and looking impressive only in a handful of titles.
Almost none of the posts here is coming from a place seeking to have an informed discussion. It's mostly fanboy propaganda that undermines an otherwise potentially very interesting conversation.
Mistershine.
Banned
I've given you a like because the word gonads is seriously underused nowadays.
Senua
Member
Last edited:
GymWolf
Member
Ps5 doesn't have it so it's shit.
geordiemp
Member
Utter nonsense.....
Show me any TAA in any game that even gets within a mile of what DLSS does.
DLSS can reconstruct a friggin 2160p image from a 1080p source and look indistinguishable from native resolution, or even better with small details.
No TAA solution in any game, ever, managed to create more than a muddy mess in that situation.
And while it is software...are you seriously denying that software with dedicated silicone runs several orders of magnitudes faster?!?
You`re announcing that no car will ever be able to do the quarter mile in under a minute while standing in front of a Ferrari. You do realize that DLSS 2.0 is already out there together with a giant number of camparison videos and analytics which all invalidate your claims?
I am not interested in a handfull of games (is it 10 ?) that have specifuic upscaling, I would prefer if a technique applies to all games bit just the ones selected for good PR.
Hopefully Ps55 will have a good mix of temporal and ML that generally can be applied to all games not just a select few.
I get why fans of team green promote the vendor specfic technique, but I really dont care.
Last edited:
Senua
Member
psorcerer's transparency is refreshing at least
Phrixotrichus
Banned
No idea what you want to tell me with a 1440p visibly "soft" looking tech demo....at least I haven`t seen any available video material where everything looked really sharp below all the PP.
You do realize that DLSS 2.0 is generic and can be used for all games without specific adaptions?
Good for you to care so much as to tell us how little you care.
Last edited:
They claim that no finetuning is needed.
Invalidates what exactly? Do we have any game that uses DNN TAA?
No? Too bad.
DLSS is an evolution of TAA.
It's a software tech.
Historically NV solutions always performed worse than game-specific ones, and I don't see why it won't be the case here too.
It's not shit.
But It's like HairWorks. Another overhyped vendor solution that will be replaced by game-specific ones in the future.
jufonuk
not tag worthy
I'm Gonna say yea Nintendo will use this or similar for a new switch pro because if it works without killing the battery and is a cheap solution to gain more power then yeah. Maybe ohh maybe it's like a super fix chip thing they did it once they can do it again.
Oh new switch pro. Has new processors etc full fat upgrade , older switch uses the chip in the carts. Little upgrade.
Oh new switch pro. Has new processors etc full fat upgrade , older switch uses the chip in the carts. Little upgrade.
Last edited:
Phrixotrichus
Banned
aaah, nothing comparable exists but you claim it will and so we`ll just have to have faith here, and believe in your fantas.....vision. Praise the lord!
"Fanboying" on a whole different level here.
You realize the hardware in the consoles will not change, right?
Once again ignoring the simple irrefutable facts at hand everyone can see by literally just booting up a game.
Last edited:
You realize that DLSS is not hardware?
Again, key points:
1. Tensor core real benchmark advantage is 2x at best over FP32 shaders.
2. All RDNA2 cards have 2x speedups for FP16 over FP32.
VFXVeteran
Banned
Yep.
Yep.
The algorithm is software true. But the processors that run this algorithm is custom. I took a deep learning class not too long ago with Nvidia and the performance converging on the solution is absolutely faster than a CPU solution for sure. Could we create an algorithm in the pixel shader units? Of course. Would that be just as fast and proficient as hardware cores that can run concurrently? Nope. AMD has essentially created a CPU inside a GPU. Not only do you implement RT using the cores, but now you have to implement DLSS in the cores as well? This is clear indication that they didn't have the tech ready to make their own hardware cores and reverted to using the basic shader cores to do all their work. The problem with that is that it soaks up bandwidth in the overall pipeline no matter what the stage. You also neglect to mention all the use cases for ML in a game. Image reconstruction isn't the only one and I can see way more complicated problems requiring vast activity from the tensor cores.
Last edited:
Phrixotrichus
Banned
And you are once again dismissing the impact of dedicated hardware.....
That´s about as pointless as saying the sky is blue in its generalization......ever visited a DL seminar? There is a LOT more to the software and the customization of the hardware.
You`re basically comparing FPGA and ASIC approaches....both exist for a reason.
Last edited:
GymWolf
Member
Lol it's not like hairworks, i barely remember anyone talking good about that thing because the perfromance hit was far greater than how it looked in game.
I see any ML as a pure software solution.
And I'm not impressed with the hw perf. On paper we have 5x-10x the flops, but in reality nobody was able to get over 2x using NV own inference libraries.
Therefore I don't think the consoles will have a disadvantage here.
Last edited:
geordiemp
Member
So UE5 demo is visibly soft, that is tacking warrior mode a step to far.
So, the list fo DLSS games, does not look generic to me
- Fortnite
- Death Stranding
- F1 2020
- Final Fantasy XV
- Anthem
- Battlefield V
- Monster Hunter: World
- Shadow of the Tomb Raider
- Metro Exodus
- Control
- Deliver Us The Moon
- Wolfenstein Youngblood
- Bright Memory
- Mechwarrior V: Mercenarie
Lets hope Ps5 patent is real and applies to all games.
Last edited:
VFXVeteran
Banned
OK. Until they show something, we can't assume they have something. We've seen several gameplay demos and not one has shown any sort of DLSS-like performance.
Phrixotrichus
Banned
You do realize that most of those games started with DLSS 1.0 which was still game-specific?
DLSS goes hand in hand with RT which is just at its starting point, too.
not at all. While the lighting and models are fantastic in the demo, not even 4k versions of the video were "crisp". This could be the compression of the video material, but considering that it`s native 1440p that seems to be the "easier" explanation.
That´s how every feature ever started, including 3d graphics.Anyway, its not important, 4000 games last gen, thats 0.3 percent....looks mainstream ....Come back when the games is 1 %
agreed.
Last edited:
Again, historically that was always the case: perf of vendor libraries (NV or AMD) was utter shit.
Right now we have nothing to compare it to. It would be nice if NV is good at software now, but I doubt it.
Yup. It would be nice if I'm wrong and an ultimate AA/upscaling solution can be made.
GymWolf
Member
Nothing to compare too?? What about every other reconstruction technique that looks worse??
They are not DL. I.e. their search space is limited.
I think the links to Medium are fucked up on neogaf, so you cannot see the article I've linked: Andrej Karpathy - Software 2.0
Dampf
Member
NSight shows that at a certain point, the tensor cores on a 2060 are to 90% used.
2060 tensor cores deliver around 53 TFLOPs of FP16 performance alone for the AI computations. For comparison, a Series X has 24 TFLOPS FP16 performance for rendering graphics and AI.
Using DLSS should have no effect to the usability of tensor cores for other tasks, as DLSS is post processing, so it is used at the very end of a frame.
Which means nothing.
Again, nobody ever was able to extract more than 2x perf from TC in Turing.
Which means the real utilization is well below 20% in any task.
Overview: https://blog.inten.to/hardware-for-deep-learning-part-3-gpu-8906c1644664
Last edited:
martino
Member
Dampf
Member
You're not wrong, but it heavily depends on the deep learning model and some benchmarks do show a different story like on page 11 here https://on-demand.gputechconf.com/ai-conference-2019/T1-1_Jack Han_Getting More DL Training with Tensor Cores and AMP_한재근_발표용.pdf
Inference should gain even more than training. Keep in mind FP32 to FP16 is not a certain 2x speedup either in most cases, there are more factors to consider than precision, like memory bandwidth.
DLSS is specifically optimized for the tensor cores, so it's fair to assume a minimum speedup of 4x and greater.
Last edited:
VFXVeteran
Banned
I think the folks that are downplaying it is because it's not shown to be available in their hardware of choice and therefore feel it is not needed. I seriously doubt they would downplay it if AMD had hardware Tensor cores. You simply wouldn't find people saying it can be easily done in software using the shader cores instead of wasting die space with tensor cores.
Last edited:
llien
Banned
2020, when people figured running at lower resolution and then upscaling improves performance.
AI inference is what you need on gaming GPUs. (And GPUs are generally great at that no matter what.)
Learning is done at data centers.
Last edited:
Zannegan
Member
You're not likely to get a non-portable Super Switch out of Nintendo unless they've been misleading their shareholders.
If they do rely on DLSS to improve their next machine's graphics, you're much more likely to get a Switch with essentially the same visuals but at apparent 1080p handheld and 2-4K docked. I doubt very much whether they'll be getting many AAA downports until their real next gen machine arrives.
If they do rely on DLSS to improve their next machine's graphics, you're much more likely to get a Switch with essentially the same visuals but at apparent 1080p handheld and 2-4K docked. I doubt very much whether they'll be getting many AAA downports until their real next gen machine arrives.
tkscz
Member
1. The PS5 and XBS use AMD. AMD has a form of software AI scaling already out but it is unknown at the moment if either console has something hardware wise
2. nVidia's DLSS is hardware based and uses the Tensor cores for Ai scaling. These tensor cores are already in handheld formats, for example the below image is the nVidia Tegra Xavier NX
That little thing is 2 -3x stronger than the switch's own nVidia Tegra X1. With 128 more CUDA cores and 42 Tensor cores that allow it to support DLSS. nVidia also says it gets 3TFLOPs with only 15w of power but I'd take that with a grain of salt. While the X1 in the Switch could do 1TFLOP at full speed (the current Switch is under clocked) I don't think the Volta cores are that much better than Maxwell (If I'm wrong, someone correct me). So the next Switch having DLSS support while still being a handheld is more than possible.
3. Nintendo went with nVidia out of desperation. The WiiU was such a huge failure, they needed something to replace it ASAP and had plans on making a hybrid handheld/console and the Tegra X1 was more than capable being that something. Sony and Microsoft already had contracts with AMD, probably long term ones from the PS4/XBO, and AMD obviously delivers on the console front, but don't really have anything like the Tegra line as far as I'm aware.
I've never seen people gaining more than 2x in inference too.
AFAIR you cannot check TC IPC in NSight. So there's no way to prove it.
alucard0712_rus
Banned
It requires motion vector data.
Omega Supreme Holopsicon
Banned
DLSS 1.9 for control ran on compute shaders without using tensor cores, iirc.
Sony patented new ai dlss like technique. They put checkerboarding h/w on the pro, it is conceivable they could have put ai upscaling h/w on the ps5. Remember not even on pro did all games make use of checkerboarding. The feature might have even been hidden on third party dev kits so microsoft didn't catch wind of it.
Last edited:
mitchman
Gold Member
Yes, a big improved over DLSS 1.0.