Quantic Dream: The shader cores of the Xbox are more suitable to machine learning

[Fafalada]

The pipeline is the pipeline.

Click to shrink...

Yes and pipeline consists of multiple frames if you look a little farther out. I think 3 separate posts pointed it out in this thread and all 3 got ignored - interleaving work with next frame has been common-place for most of this decade (give or take) and if you have an 'ideal' scenario where hypothetical hw-runs completely independent (as NVidia claimed) - you get 100% parallelism out of it, which would make the cost of ML portion for DLSS equal to 0. There's also a number of ways how that could work within single frame - but the point is it doesn't really have to.
The contended point is apparently NVidia wasn't entirely honest on what the hw can actually do, or at least so it seems.

And you cant apply post processing without first upscaling the frame.

Click to shrink...

Of course you can. And many, many reconstructed games do exactly that, a lot (sometimes all) of their PP runs before reconstruction, not after. And we see evidence of that in DLSS games as well.

The public doesnt really care about how efficient the use of the silicon is.

Click to shrink...

This isn't really efficiency question so much as misleading the public (if tensor cores indeed can only execute serially to the rest of GPU).

 

[Alucardx23]

If the consoles benefit at all from ML, it doesn't seem like it will be for framerate improvements. Which would be an unfortunate lategen turn, but we're already seeing it with rtx, so.

Click to shrink...

I don't know about that, a lot of rendering time can be saved by not having to render a native 4K frame VS a 1080P one. The XSX has half the TOPs performance than a RTX 2060. With that we can estimate that to do a DLSS equivalent step at the same quality, it will take double the amount of time than the 2060 does. The thing is that it can use a lower quality ML image upscaling technique or it can be a more or optimized solution than can do more with less. That is something that we need to wait and see.

 

[Alucardx23]

Alright, maybe we're talking past each other. Let's just agree to disagree on the point being made. I've (and some others pm'ing me) said what we are trying to say and it obviously is not being received.

Let's just let the performance and quality of AMD's upcoming solution do the talking.

Click to shrink...

I don't know where we are talking past each other. You said that the ML upscaling step will take away from the traditional rendering step and so far it all points to the ML upscaling step needing to have a complete rendered frame, prior to the post processing steps. I have shared graphs and videos from Nvidia explaining how this works and direct quotes from reliable sources. All very clearly talk about how DLSS does not work concurrently with the ALU. Ask yourself if you have provided equivalent evidence to prove your point that they are in fact concurrent? A simple yes or no answer will do.

 

[Muppet of a Man]

I don't know where we are talking past each other. You said that the ML upscaling step will take away from the traditional rendering step and so far it all points to the ML upscaling step needing to have a complete rendered frame, prior to the post processing steps. I have shared graphs and videos from Nvidia explaining how this works and direct quotes from reliable sources. All very clearly talk about how DLSS does not work concurrently with the ALU. Ask yourself if you have provided equivalent evidence to prove your point that they are in fact concurrent? A simple yes or no answer will do.

Click to shrink...

I will answer by quoting another poster because I really don't have time to keep up with this thread at the moment:

Okay, let's agree that ML reconstruction doesn't work concurrently with CU shader operations. Now, imagine a game that already uses DLSS removes it. Does the GPU now sit completely idle during the time that DLSS used to run on tensor cores? Of course not, it runs CU shader ops during that time.

Now add DLSS back in. Those CU ops can't be done anymore, because the time is taken up by ML upscaling ops. In other words, you are robbing the GPU of cores for actual traditional rendering tasks. The proper conclusion of your argument isn't "it's not a choice between ML upscale and typical rendering", it's "you are always trading away some typical rendering to do ML upscale, even on GPUs with dedicated tensor cores".


That quote's correct interpretation hinges on what was meant by "any ML stuff". Did they mean any hardware at all that could be useful for ML tasks? Or did they mean dedicated cores like those used for ML elsewhere? I can definitely see the plausibility of the former reading. But the immediate following mention of a different kind of dedicated core could point to the latter meaning.

I personally think PS5 likely doesn't have extra hardware for small INT packing. But we don't know for sure.

Click to shrink...

 

[Alucardx23]

I will answer by quoting another poster because I really don't have time to keep up with this thread at the moment:

Click to shrink...

Okay, let's agree that ML reconstruction doesn't work concurrently with CU shader operations. Now, imagine a game that already uses DLSS removes it. Does the GPU now sit completely idle during the time that DLSS used to run on tensor cores? Of course not, it runs CU shader ops during that time.

Now add DLSS back in. Those CU ops can't be done anymore, because the time is taken up by ML upscaling ops. In other words, you are robbing the GPU of cores for actual traditional rendering tasks. The proper conclusion of your argument isn't "it's not a choice between ML upscale and typical rendering", it's "you are always trading away some typical rendering to do ML upscale, even on GPUs with dedicated tensor cores".

Click to shrink...

See the amount of times I have said that what you just said doesn't have to do anything at all with my point. You could ask the same question with the RTX GPUs. I mean, why not give the 1.5ms the tensor cores takes to do the DLSS step back to the cuda cores? That question doesn't make any sense right? If you don't do the DLSS step, then simple you have a 1080P image instead of a ML 4K one. The whole point of doing the ML image upscaling step is to render the image at a lower resolution, to then do the upscaling to something that resembles 4K and gain performance doing so. You don't get to a ML upscaled 4K image by giving that 1.5ms time back to the cuda cores. Who in their right mind would think that the debate here is between people that say that the ML Upcaling step takes time on the XSX CU VS the ones that say that it doesn't take any time? Of course it takes GPU time. The corrections here have been related to how at least on the RTX GPUs the DLSS does not run concurrently with the cuda cores and how in general that is how ML image upscaling techniques seem to work. If this same process is followed on the XSX it would not be robbing anything out of the CU on the XSX, because the traditional rendering step would be completed before the ML upscaling step starts and that is the key part here. They are not concurrent steps.

That quote's correct interpretation hinges on what was meant by "any ML stuff". Did they mean any hardware at all that could be useful for ML tasks? Or did they mean dedicated cores like those used for ML elsewhere? I can definitely see the plausibility of the former reading. But the immediate following mention of a different kind of dedicated core could point to the latter meaning.

I personally think PS5 likely doesn't have extra hardware for small INT packing. But we don't know for sure.

Click to shrink...

Of course we need more details on that, but it can be related to how the CU on the PS5 not being enabled for INT4 and INT8. We will need to wait and see of that.

 

[Liabe Brave]

The corrections here have been related to how at least on the RTX GPUs the DLSS does not run concurrently with the cuda cores and how in general that is how ML image upscaling techniques seem to work.

Click to shrink...

Well, the Beyond3D post linked earlier says otherwise, as does a multiplat dev posting here. But even if this is inaccurate speculation, and your reading is correct, your conclusion still doesn't follow. See below.

If this same process is followed on the XSX it would not be robbing anything out of the CU on the XSX, because the traditional rendering step would be completed before the ML upscaling step starts and that is the key part here. They are not concurrent steps.

Click to shrink...

And what other people have tried to explain to you is that whether the work is concurrent or sequential has no bearing on total resources needed. Either way, unless you add compute to do the extra step(s), you are giving up standard compute to achieve ML upscaling.

To state the point another way: DLSS obviously has to finish before the frame it's working on is needed to display. Let's say we have a 60fps game, and DLSS takes ~1.5ms. That means normal rendering has to be done within ~15.1ms. But what if DLSS is off? Now normal rendering has 16.6ms to work in. In other words, you can do more normal rendering now.

Or maybe a reductio will help. You say that because normal rendering is finished when DLSS starts, you don't get any resources back by turning DLSS off. But what about games without DLSS? Their normal rendering is finished before, say, chromatic aberration starts. Do you get any resources back by turning CA off? Maybe not, but now all normal rendering is finished before temporal antialiasing starts. Do you get any resources back turning TAA off? And now the last step is per-pixel motion blur. Do you get any resources back turning it off? Now all normal rendering is done before particle emission. Remove them, and all normal rendering is done before raytracing. Does removing RT not get you any resources back?

In other words, as you start turning them off, every penultimate step becomes the last step. If shutting off the last step frees up nothing, eventually you can shut off every feature...but supposedly still not free up any resources?

 

[NinjaGarden]

Xbox Series X is more likely to gain sentience and turn on you. Beware.

 

[mordecaii83]

The contended point is apparently NVidia wasn't entirely honest on what the hw can actually do, or at least so it seems.

Click to shrink...

Just out of curiosity, where is the claim that tensor cores can't run alongside CUDA cores/RTX cores concurrently coming from? I haven't been following this thread closely, but the whitepaper for Ampere seems to state all 3 can run concurrently and the same whitepaper states this is a change from Turing in which they could not.

Edit: I was mistaken, Turing couldn't run RT + Shading/Compute at the same time, Ampere can. Both seem to show concurrent Tensor/Shading though from what I can tell on their whitepapers?

 

[Trup1aya]

Yes and pipeline consists of multiple frames if you look a little farther out. I think 3 separate posts pointed it out in this thread and all 3 got ignored - interleaving work with next frame has been common-place for most of this decade (give or take) and if you have an 'ideal' scenario where hypothetical hw-runs completely independent (as NVidia claimed) - you get 100% parallelism out of it, which would make the cost of ML portion for DLSS equal to 0. There's also a number of ways how that could work within single frame - but the point is it doesn't really have to.
The contended point is apparently NVidia wasn't entirely honest on what the hw can actually do, or at least so it seems.


Of course you can. And many, many reconstructed games do exactly that, a lot (sometimes all) of their PP runs before reconstruction, not after. And we see evidence of that in DLSS games as well.


This isn't really efficiency question so much as misleading the public (if tensor cores indeed can only execute serially to the rest of GPU).

Click to shrink...

The public hasn't been mislead, Nvidia has never claimed, as far as I can tell, that these are parallel processes. In fact they've explicitly said they arent in the demonstrations I've seen.

It seems to me that people have made their own determinations on the benefit of "dedicated ML hardware" being separate from the CU, when really the benefit is nvidia has more space for more hardware so that upscaling can be peformed faster.

The pipeline consisting of multiple frames doesnt mean that the rendering of these frames his happening in parallel with the upscaling process. However, since Nvidia's cards can upscale faster thanks to having more hardware dedicated to upscaling, they can 1) more quickly begin rendering the next frame, or 2) they can spend more time post-processing the current frame, or 3) they can spend more frametime creating a higher resolution native image.

 

[Trup1aya]

Well, the Beyond3D post linked earlier says otherwise, as does a multiplat dev posting here. But even if this is inaccurate speculation, and your reading is correct, your conclusion still doesn't follow. See below.


And what other people have tried to explain to you is that whether the work is concurrent or sequential has no bearing on total resources needed. Either way, unless you add compute to do the extra step(s), you are giving up standard compute to achieve ML upscaling.

To state the point another way: DLSS obviously has to finish before the frame it's working on is needed to display. Let's say we have a 60fps game, and DLSS takes ~1.5ms. That means normal rendering has to be done within ~15.1ms. But what if DLSS is off? Now normal rendering has 16.6ms to work in. In other words, you can do more normal rendering now.

Or maybe a reductio will help. You say that because normal rendering is finished when DLSS starts, you don't get any resources back by turning DLSS off. But what about games without DLSS? Their normal rendering is finished before, say, chromatic aberration starts. Do you get any resources back by turning CA off? Maybe not, but now all normal rendering is finished before temporal antialiasing starts. Do you get any resources back turning TAA off? And now the last step is per-pixel motion blur. Do you get any resources back turning it off? Now all normal rendering is done before particle emission. Remove them, and all normal rendering is done before raytracing. Does removing RT not get you any resources back?

In other words, as you start turning them off, every penultimate step becomes the last step. If shutting off the last step frees up nothing, eventually you can shut off every feature...but supposedly still not free up any resources?

Click to shrink...

I'm not following this at all.
Whether you use AMDs method or Nvidia's method, "turning DLSS off" would mean more time for normal rendering because you are no longer dedicating any frametime to upscaling.

On the Nvidia card, during the 1.5ms of DLSS, the GPU isnt doing anything. You turn off DLSS, now the GPU has the full 16.6ms to work in.

On AMDs card, we may need 5.5ms of DLSS because we have less dedicated hardware compared to the nvidia card - So just 11.1ms for "normal rendering". But if we cut DLSS off, then we're right back at having 16.6ms to work in.

 

[melodiousmowl]

I'm not following this at all.
Whether you use AMDs method or Nvidia's method, "turning DLSS off" would mean more time for normal rendering because you are no longer dedicating any frametime to upscaling.

On the Nvidia card, during the 1.5ms of DLSS, the GPU isnt doing anything. You turn off DLSS, now the GPU has the full 16.6ms to work in.

On AMDs card, we may need 3ms of DLSS because we have less dedicated hardware compared to the nvidia card - So just 13.6ms for "normal rendering". But if we cut DLSS off, then we're right back at having 16.6ms to work in.

Click to shrink...

what does a frame, all things but resolution being the same, take to render at different resolutions?

is the time delta > the time it take to dlss the frame?

i have to assume it is, especially as the graphic effects go higher

 

[Trup1aya]

what does a frame, all things but resolution being the same, take to render at different resolutions?

is the time delta > the time it take to dlss the frame?

i have to assume it is, especially as the graphic effects go higher

Click to shrink...

It's much faster to "DLSS the frame" than to render the frame natively in 4k. That saved time is then used by the GPU further beautify the image.

AMDs chip has the TOPs to work at half the speed of the the RTX2060S. If AMD could run the same algorithm and extrapolating linearly, the cost of getting a 1080p image to 4k would be 5.5s on the XSX.

 

[Kaldaien]

Xbox adhering to the standard would make it better for generic GPGPU workloads

Click to shrink...

Which make more sense on some kind of backend / development toolset than in the actual console games are played on. A GPU that's good at machine learning is one that belongs in the developer's hands crunching training sets, the actual delivered product doesn't need strong general purpose compute capabilities.

 

[Kaldaien]

Forgive me if I've got this wrong, but isn't the Xbox algorithm for auto hdr, also based on machine learning (or something similar) ?

Click to shrink...

Yes, but the training is all done on Microsoft's end. The console itself isn't learning anything to derive the tonemaps.

 

[Kaldaien]

Those aren't GPUs.

Colloquially the fastest GPU has always meant the most powerful. I'm not sure why you are convinced otherwise.

Click to shrink...

Way back in the day, the fastest GPU was the one with the fastest RAMDAC, not the one that filled the most pixels in a unit time :) You paid big bucks for high resolution / refresh rates, and didn't always have a higher pixel fillrate to match. This is what separated workstation GPUs from consumer ones.

 

[harz-marz]

I understand none of this but still find it fascinating to read!

 

[Alucardx23]

Well, the Beyond3D post linked earlier says otherwise, as does a multiplat dev posting here. But even if this is inaccurate speculation, and your reading is correct, your conclusion still doesn't follow. See below.

Click to shrink...

You mean the Beyond 3D post below that clearly states how the DLSS step has to wait for the frame to be completed? This is about how DLSS works and what we read below is almost exactly what I'm saying. Nevermind all of the other information I have shared that probes this.

"DLSS replaces the MSAA step that would normally be handled by your GPU. During this time it will DLSS and the final results will be spit back for the rest of the pipeline to continue. When Turing/Ampere offloads the work away from the CUDA cores to Tensor Cores to do this, the pipeline, as I understand from reading, is stalled until DLSS is complete. So you can't really continue rendering normally until DLSS is done. You can (edit: maybe not) do async compute while waiting for DLSS to complete since your compute is free to do whatever as you offloaded this process to Tensor Cores."

To state the point another way: DLSS obviously has to finish before the frame it's working on is needed to display. Let's say we have a 60fps game, and DLSS takes ~1.5ms. That means normal rendering has to be done within ~15.1ms. But what if DLSS is off? Now normal rendering has 16.6ms to work in. In other words, you can do more normal rendering now.

Or maybe a reductio will help. You say that because normal rendering is finished when DLSS starts, you don't get any resources back by turning DLSS off.

Click to shrink...

The whole point of doing the ML image upscaling step is to render the image at a lower resolution, to then do the upscaling to something that resembles 4K and gain performance doing so. You don't get to a ML upscaled 4K image by giving that 1.5ms time back to the cuda cores. Who in their right mind would think that the debate here is between people that say that the ML Upcaling step takes time on the XSX CU VS the ones that say that it doesn't take any time? Of course it takes GPU time. The corrections here have been related to how at least on the RTX GPUs the DLSS does not run concurrently with the cuda cores and how in general that is how ML image upscaling techniques seem to work. If this same process is followed on the XSX it would not be robbing anything out of the CU on the XSX, because the traditional rendering step would be completed before the ML upscaling step starts and that is the key part here. They are not concurrent steps.

Click to shrink...

I'm done with you and your strawmans. It should be more than clear from the posts above that you are either too lazy to read the complete post or are willingly misinterpreting what the other person is saying. Either way it's a waste of time. Good luck with the strawmans tactics.

 

[arsene_P5]

Yep, there's clearly a huge bias on Era towards Xbox.

Click to shrink...

Sure xD

Fanboys want their 'gotcha' moment. Till Sony releases public info on their APU it's all faceless speculation.

Click to shrink...

What gotcha moment? Both consoles have customizations and this according to him is one Xbox has a advantage on. If you believe Sony will come out and say something about it, then you are probably in for a lengthy wait. Both companies won't comment on the other products benefits and tell their consumers, they don't have them.

I think the reaction to his opinion about the Series S differ, because there are also developers saying that scaling for different targets is a non issue, if you develop for PC and are experienced with this and devs who were using the more powerful console as the lead plattform in the past. So his opinion can be argued against, while with this ML statement we don't know any better as consumers and there is no reason for him to lie.

Just to clarify, I don't think he lied with his Series S statement. He just gave his opinion and his concern, which some devs agree with and others don't.

Edit:
I recall someone asking if we know who is planning to use these ML stuff. Here you go and btw Ninja Theory is rumoured to use these techniques for one of their projects, too.

As a Senior Machine Learning Engineer at The Coalition, you will be a highly valued member of a world class Engine and Game Development team. You should have a passion for applying machine learning techniques to innovate and advance game development and title features across a broad set of areas such as rendering, data science, compression, and artificial intelligence.

Click to shrink...

https://careers.microsoft.com/us/en/job/926910/Senior-Machine-Learning-Engineer-The-Coalition

 

[Liabe Brave]

I'm not following this at all.
Whether you use AMDs method or Nvidia's method, "turning DLSS off" would mean more time for normal rendering because you are no longer dedicating any frametime to upscaling.

Click to shrink...

I totally agree. But Alucardx23 apparently both does and does not, since he's made statements both ways. This is why I'm debating with him.

AMDs chip has the TOPs to work at half the speed of the the RTX2060S. If AMD could run the same algorithm and extrapolating linearly, the cost of getting a 1080p image to 4k would be 5.5ms on the XSX.

Click to shrink...

If that estimate is approximately correct, you can basically count out DLSS at 60fps for all but the simplest games. Most titles will be unwilling to spend a third of their rendering budget on reconstruction. Even at 30fps a 5.5ms task is pretty heavy, but it at least doesn't seem prohibitive for some games.

You mean the Beyond 3D post below that clearly states how the DLSS step has to wait for the frame to be completed?

Click to shrink...

It does also bring up the possibility of async compute. But as I said, the potential inaccuracy of this post has no effect on what I'm saying.

I'm done with you and your strawmans. It should be more than clear from the posts abode that you are either too lazy to read the complete post or are willingly misinterpreting what the other person is saying.

Click to shrink...

I assure you that what I wrote is my true understanding of what you've been saying. If it's inaccurate, I apologize. If you're willing, feel free to point out what I've gotten wrong here. Here's the key quote of yours:

You said "if you run ML-trained algorithms on CUs, even if they utilized INT8 or INT4, you are robbing the GPU of cores for actual traditional rendering tasks.". I just shared the information that shows that the image reconstruction step has to wait for the frame to be complete in order to start working on it. It does not run concurrently with the traditional rendering tasks, so no robbing happens when doing image upscaling using ML.

Click to shrink...

The claim being debated is that running DLSS reduces the compute power for standard rendering. You reject that, and your evidence is that DLSS only takes place after standard rendering is complete. Therefore, you conclude, no reduction in power occurs.

I don't see any other reasonable way to interpret your meaning. Yes, you've stated that you do agree that DLSS takes time (of course). This is exactly why it's so weird for you to simultaneously assert that "no robbing [of GPU power] happens". The inconsistency is why multiple people have debated with you.

 

[Pyro]

Isn't this kind of a "no shit" thing? We've known for a while that Series X would have dedicated hardware for machine learning no?

 

[Don Dada]

Fam what is being argued here. Only one fact we need to know, DLSS and similar tech means that a GPU can put more sexy shit on screen rendering at a lower than native resolution whilst making it looks as good if not better than native resolution.

If a GPU supports it then we get more sexy shit on the screen. It doesn't matter if it renders it before after or during traditional tasks.

Everything has a opportunity cost and DLSS Is more than worth it.

 

[Negotiator117]

Isn't this kind of a "no shit" thing? We've known for a while that Series X would have dedicated hardware for machine learning no?

Click to shrink...

It was mentioned in the Digital Foundry reveal but everyone presumed the PS5 had the same hardware modifications, now we know this is an Xbox exclusive.

 

[Pyro]

It was mentioned in the Digital Foundry reveal but everyone presumed the PS5 had the same hardware modifications, now we know this is an Xbox exclusive.

Click to shrink...

Another thing I thought was also known. Or maybe not known, but I didn't expect PS5 to have it.

 

[Negotiator117]

Another thing I thought was also known. Or maybe not known, but I didn't expect PS5 to have it.

Click to shrink...

It's not something I expect to come in to play till later on in the generation, unless we get a mid gen upgrade but I don't expect that this time.

 

[Trup1aya]

I totally agree. But Alucardx23 apparently both does and does not, since he's made statements both ways. This is why I'm debating with him.


If that estimate is approximately correct, you can basically count out DLSS at 60fps for all but the simplest games. Most titles will be unwilling to spend a third of their rendering budget on reconstruction. Even at 30fps a 5.5ms task is pretty heavy, but it at least doesn't seem prohibitive for some games.


It does also bring up the possibility of async compute. But as I said, the potential inaccuracy of this post has no effect on what I'm saying.


I assure you that what I wrote is my true understanding of what you've been saying. If it's inaccurate, I apologize. If you're willing, feel free to point out what I've gotten wrong here. Here's the key quote of yours:



The claim being debated is that running DLSS reduces the compute power for standard rendering. You reject that, and your evidence is that DLSS only takes place after standard rendering is complete. Therefore, you conclude, no reduction in power occurs.

I don't see any other reasonable way to interpret your meaning. Yes, you've stated that you do agree that DLSS takes time (of course). This is exactly why it's so weird for you to simultaneously assert that "no robbing [of GPU power] happens". The inconsistency is why multiple people have debated with you.

Click to shrink...

I'm butting in late but I've kinda been following the debate and it seemed me that people we arguing that with Nvidia's approach, the GPU isnt losing time to DLSS because it has dedicated hardware. But in reality, both approaches have "dedicated hardware", and it both cases the DLSS processes uses frametime that would other wise be used for "normal rendering". Alucard wasn't saying that on AMD there is no cost to DLSS on the GPU. He was saying that on Nvidia, the GPU loses frametime at the same step in the process... and having tensor cores doesnt alleviate this. I feel like Alucardx23 was the one who WASN'T trying to have it both ways.

Nvidia's cards simply have more dedicated hardware than what can be squeezed into the CUs of a console GPU... so the GPU has more time to work.

As far as framerates on XSX using a DLSS-like approach, I'd wager that 5.5ms to DLSS from 1080p to 4K would allow for more time "normal rendering" a 4k60 game than we'd get on a game that is native 4k60. But your right, I would expect a lot of 4k30 with this approach on a device with 49TOPs..

 

[Jeepman87]

Forgive me if I've got this wrong, but isn't the Xbox algorithm for auto hdr, also based on machine learning (or something similar) ?

Click to shrink...

Yes

 

[ItIsOkBro]

To state the point another way: DLSS obviously has to finish before the frame it's working on is needed to display. Let's say we have a 60fps game, and DLSS takes ~1.5ms. That means normal rendering has to be done within ~15.1ms. But what if DLSS is off? Now normal rendering has 16.6ms to work in. In other words, you can do more normal rendering now.

Click to shrink...

I don't get this. "Normal rendering" would take less time with DLSS on, that's the whole point ain't it.

 

[Alucardx23]

I'm butting in late but I've kinda been following the debate and it seemed me that people we arguing that with Nvidia's approach, the GPU isnt losing time to DLSS because it has dedicated hardware. But in reality, both approaches have "dedicated hardware", and it both cases the DLSS processes uses frametime that would other wise be used for "normal rendering". Alucard wasn't saying that on AMD there is no cost to DLSS on the GPU. He was saying that on Nvidia, the GPU loses frametime at the same step in the process... and having tensor cores doesnt alleviate this. I feel like Alucardx23 was the one who WASN'T trying to have it both ways.

Nvidia's cards simply have more dedicated hardware than what can be squeezed into the CUs of a console GPU... so the GPU has more time to work.

As far as framerates on XSX using a DLSS-like approach, I'd wager that 5.5ms to DLSS from 1080p to 4K would allow for more time "normal rendering" a 4k60 game than we'd get on a game that is native 4k60. But your right, I would expect a lot of 4k30 with this approach on a device with 49TOPs..

Click to shrink...

Thank you for steelmanning my position.

 

[Alucardx23]

I don't get this. "Normal rendering" would take less time with DLSS on, that's the whole point ain't it.

Click to shrink...

Exactly, that is the whole point, to save performance and get to a close enough image quality to native. Trying to do the same thing without the ML upscaling step, would reduce the performance, that is the whole reason why you would use ML upscaling. When you turn on DLSS on any PC game that supports it, you get to increase the graphics settings and framerate more than without having DLSS. Developers get to be more efficient with the hardware and still get to the resolution they want.

 

[Liabe Brave]

Fam what is being argued here. Only one fact we need to know, DLSS and similar tech means that a GPU can put more sexy shit on screen rendering at a lower than native resolution whilst making it looks as good if not better than native resolution.

Click to shrink...

DLSS does not look as good or better than native resolution in general, though Nvidia and Digital Foundry have said or implied this multiple times. Like all other types of reconstruction, certain elements can be equivalent in this way--for DLSS specifically, subpixel detail is very well recovered for example. But also like all other types of reconstruction, most elements are less accurate than native rendering. This is why DLSS includes a very strong sharpening filter, because its raw results are considerably blurrier than native. It also generates lots of temporal artifacts.

This isn't to say that DLSS is pointless or bad. See below for more.

I'm butting in late but I've kinda been following the debate and it seemed me that people we arguing that with Nvidia's approach, the GPU isnt losing time to DLSS because it has dedicated hardware. But in reality, both approaches have "dedicated hardware", and it both cases the DLSS processes uses frametime that would other wise be used for "normal rendering". Alucard wasn't saying that on AMD there is no cost to DLSS on the GPU. He was saying that on Nvidia, the GPU loses frametime at the same step in the process... and having tensor cores doesnt alleviate this.

Click to shrink...

This all sounds like an accurate paraphrase to me...but it isn't the whole of his position. In addition to all the above, he has also been repeatedly stated that, because DLSS happens after standard rendering, saying that it reduces GPU resources for standard rendering is false. This part is false (and doesn't fit with his other statements you've described), regardless of how DLSS operates.

I don't get this. "Normal rendering" would take less time with DLSS on, that's the whole point ain't it.

Click to shrink...

Rendering at much lower resolution is less demanding, yes. But the point was merely that there would be even more time to devote to rendering, at whatever resolution, if DLSS wasn't included.

Exactly, that is the whole point, to save performance and get to a close enough image quality to native. Trying to do the same thing without the ML upscaling step, would reduce the performance, that is the whole reason why you would use ML upscaling.

Click to shrink...

Of course, and the same is true for other reconstruction methods as well. CBR, temporal injection, etc. also allow the same tradeoff, reducing standard rendering load by lowering target resolution, and completing the "missing" pixels through a much faster, though also less accurate, method.

 

[Don Dada]

I don't know man from the DF article for death stranding

Comparisons between the two techniques are fascinating but the big takeaway is that DLSS image reconstruction from 1440p looks cleaner overall than native resolution rendering. We've reached the point where upscaling is quantifiably cleaner and more detailed - which sounds absurd, but there is an explanation.

That's enough for me.