• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

Cerny confirms PSSR2 and FSR4.1 use same neural network

winjer said:
No. It's done using a mix of data formats. Including Integers.
But even after training, the resulting model can be quantitized into lower precision models.
Click to expand...
No... all training is done in FP. You can't train in int8 because of a values limitation, the training model simply will not work. So even when training for an int execution unit, they use FP that is "pretending" to be int. FP preends to be Int by adding fake noise into the training model to account for the rounding errors that are unavoidable with INT.

Once the training is done however, it can be compressed down to INT... but it will never be as accurate as a model running in FP. This could explain why even as great as PSSR2 is, there is a tell with how it handles super fine repetitiuve detail with that weird pointilism thing it does.
 
Mr.Phoenix said:
No... all training is done in FP. You can't train in int8 because of a values limitation, the training model simply will not work. So even when training for an int execution unit, they use FP that is "pretending" to be int. FP preends to be Int by adding fake noise into the training model to account for the rounding errors that are unavoidable with INT.

Once the training is done however, it can be compressed down to INT... but it will never be as accurate as a model running in FP. This could explain why even as great as PSSR2 is, there is a tell with how it handles super fine repetitiuve detail with that weird pointilism thing it does.
Click to expand...

Int is also used, although in limited scale. Such as with in Quantization-Aware Training.

Still, quantitizing a trained model to Int8, is the easy part of the process.

Besides, I never said that training was done in Int8.
 
Last edited:
winjer said:
Int is also used, although in limited scale. Such as with in Quantization-Aware Training.

Still, quantitizing a trained model to Int8, is the easy part of the process.
Click to expand...
I believe QAT is when training is done with FP that pretends to be INT by adding fake noise into the model. INT is mathematically incapable of being used for training a model, thats why training is done in FP. QAT trains a model using FP that is already accounting for the limitations of INT, its still technically an FP output, just rounded down for INT. The great thing is that the same output can run natively in FP too. In its pre-rounded form of course.

This is likely how sony and AMD's models are trained. Trained using QAT, rounded down and exported for INT while the OG (before rounding) value is ran natively in FP. This is how they can train one model that caters to both platforms. And even with PSSR3, sony is likely sticking with INT, for all Ai functions except those that explicitly need FP.
 
Last edited:
Mr.Phoenix said:
I believe QAT is when training is done with FP that pretends to be INT by adding fake noise into the model. INT is mathematically incapable of being used for training a model, thats why training is done in FP. QAT trains a model using FP that is already accounting for the limitations of INT, its still technically an FP output, just rounded down for INT. The great thing is that the same output can run natively in FP too.

This is likely how sony and AMD's models are trained. Trained using QAT, rounded down and exported for INT while the OG (before rounding) value is ran natively in FP. This is how they can train one model that caters to both platforms. And even with PSSR3, sony is likely sticking with INT, for all Ai functions except those that explicitly need FP.
Click to expand...

We are saying the same thing essentially.
But I never said training was done in Int8. I don't even understand why you suddenly started this line of argument.
 
winjer said:
We are saying the same thing essentially.
But I never said training was done in Int8. I don't even understand why you suddenly started this line of argument.
Click to expand...
My bad, you said something about converting from FP to INT not being the problem and that what was, was the training. I thought you were saying that its training in INT that was a problem. I also felt that training in general isn't ven a problem because of QAT, which is most likely what is being used here by both sony and AMD because it kills two biirds with one stone.
 
winjer said:
We are saying the same thing essentially.
But I never said training was done in Int8. I don't even understand why you suddenly started this line of argument.
Click to expand...
Because you claimed the training is the hard time consuming part making it seem like they had to do something different. The argument didn't make sense.
 
Mr.Phoenix said:
My bad, you said something about converting from FP to INT not being the problem and that what was, was the training. I thought you were saying that its training in INT that was a problem.
Click to expand...

My reference was just that training a model takes a lot more time and effort, than to quantitize it to different data formats, such as FP8 and Int8. I never said anything about which data formats are used in training.

Mr.Phoenix said:
I also felt that training in general isn't ven a problem because of QAT, which is most likely what is being used here by both sony and AMD because it kills two biirds with one stone.
Click to expand...

AMD and Sony probably had to develop special techniques in a joint effort to be able to easily and quatitize these models into either RDNA4 and the Pro.
 
Three said:
Because you claimed the training is the hard time consuming part making it seem like they had to do something different. The argument didn't make sense.
Click to expand...

That is because training a model really is the hardest part of the process. Converting the model to a lower data format is the easy part of the whole process.
 
winjer said:
That is because training a model really is the hardest part of the process. Converting the model to a lower data format is the easy part of the whole process.
Click to expand...
Sure but why did you bring it up as if it proved that the reason FSR4 didn't come to PS5 Pro but FSR4.1 did was due to training?

Your argument didn't make sense at all. People inferred that you were claiming they had to train in int8 for the Pro.

You completely dodged my post asking you this too. You still haven't explained how the easy part (not the training you claimed was time consuming) prevented FSR4 on PS5 Pro happening earlier.
 
Last edited:
Three said:
Sure but why did you bring it up as if it proved that the reason FSR4 didn't come to PS5 Pro but FSR4.1 did was due to training?

Your argument didn't make sense at all. People inferred that you were claiming they had to train in int8 for the Pro.

You completely and dodged my post asking you this too. You still haven't explained how the easy part (not the training you claimed was time consuming) prevented FSR4 on PS5 Pro happening earlier.
Click to expand...

Because FSR4 was designed with and for AMD's software stack. RoCm and Hip.
It targeted the performance metrics of RDNA4 and it's ML units.
And it's quantitization process was developed for FP8.
It also has to take into account the OS and API. Which are very different. Just consider that AMD still has not made a version of FSR4 for Vulkan.
So it took time to develop a common software stack, and quantitization processes.
If Sony was making FSR4 from the very start, with the goal of putting it on the Pro, then FSR4.0 would be there a year ago. Instead it's here now, with FSR4.1
 
Last edited:
winjer said:
Because FSR4 was designed with and for AMD's software stack. RoCm and Hip.
It targeted the performance metrics of RDNA4 and it's ML units.
And it's quantitization process was developed for FP8.
It also has to take into account the OS and API. Which are very different.
So it took time to develop a common software stack, and quantitization processes.
If Sony was making FSR4 from the very start, with the goal of putting it on the Pro, then FSR4.0 would be there a year ago. Instead it's here now, with FSR4.1
Click to expand...

I'm not exactly sure, why are you talking about this

They skipped 4.0 and decided to go straight to a "4.1" equivalent....

Who cares at this point?
 
winjer said:
And it's quantitization process was developed for FP8.
It also has to take into account the OS and API. Which are very different. Just consider that AMD still has not made a version of FSR4 for Vulkan.
So it took time to develop a common software stack, and quantitization processes.
If Sony was making FSR4 from the very start, with the goal of putting it on the Pro, then FSR4.0 would be there a year ago. Instead it's here now, with FSR4.1
Click to expand...
But this was exactly what you were dismissing as the reason for that delay:

Three said:
Has nothing to do with it. You clearly said

Blatantly incorrect. These things do take time to develop which was why PSSR was launched earlier. They continued working on the next iteration which was FSR4 but the algorithm was fp8 while the delayed PS5 Pro was int8. That took time and effort to implement and AMD did not even bother doing so for their older PC GPUs. Doesn't mean PSSR and FSR4 were separate models before though.

Click to expand...
And you replied:
winjer said:
Converting a model from Int8 to FP8 and vice versa is not the thing that takes more time. It's the training.
Saying that Sony was using PSSR1 for a year, while having made FSR4, makes no sense at all.
Click to expand...
They had trained FSR4 already. Why were you bringing up the training if the next step of quantitization, API/OS, etc was the easy part? That's why people inferred that you're trying to suggest there was some additional time consuming training specifically for int8/PS5 Pro.

The argument didn't make sense. You then continued to dismiss AMD and Cerny tweets explicitly stating the joint development of FSR4.
 
Last edited:
Three said:
But this was exactly what you were dismissing as the reason for that delay:

And you replied:

They had trained FSR4 already. Why were you bringing up the training if the next step of API/OS, etc was the easy part? That's why people inferred that you're trying to suggest there was some additional time consuming training specifically for int8/PS5 Pro.

The argument didn't make sense. You then continued to dismiss AMD and Cerny tweets explicitly stating the joint development of FSR4.
Click to expand...

You don't just develop a model. It requires a software stack, both for development and for running the model.
AMD started developing FSR4 targeting RDNA4, on Windows, in DirectX, to run on FP8 quatitization. And that is what came out of FSR4.0
When Sony entered the picture, a lot had to be added and/or changed in the process. And it took time, until we got FSR4.1
 
winjer said:
You don't just develop a model. It requires a software stack, both for development and for running the model.
AMD started developing FSR4 targeting RDNA4, on Windows, in DirectX, to run on FP8 quatitization. And that is what came out of FSR4.0
When Sony entered the picture, a lot had to be added and/or changed in the process. And it took time, until we got FSR4.1
Click to expand...
Training doesn't rely on windows, DirectX or int8. It was the "running the model" on PS5 Pro that you were refuting and claiming was easy to do despite it being int8 and RDNA2, claiming it didn't explain the time it took to come to Pro but suggesting training did. What you were saying made no logical sense, sorry.
 
Last edited:
Three said:
Training doesn't rely on windows, DirectX or int8. It was the "running the model" on PS5 Pro that you were refuting and claiming was easy to do despite it being int8 and RDNA2, claiming it didn't explain the time it took to come to Pro but training did. What you were saying made no logical sense, sorry.
Click to expand...

OMG, you can't be serious. Do you really think a model just runs on itself by magic?
 
KeplerL2 said:
FSR 5/Diamond are transformer models for upscaling/framegen/denoising built for RDNA5. Will be pretty much 100% identical between PS6/Xbox Helix/RDNA5 dGPUs regardless of whatever branding each one uses.
Click to expand...

Oh WOW! Okay, I like where AMD is headed now.

Closer said:
Maybe Sony made a deal to have the only INT8 version avaliable, who knows.
Click to expand...

Never considered this. But this sounds interesting.
 
winjer said:
OMG, you can't be serious. Do you really think a model just runs on itself by magic?
Click to expand...
What does that even mean? A trained model is not really dependant on "DirectX and windows". It's running that trained model on different hardware/OS/int8 that requires work but you dismissed that time consuming work to mention 'training' as the issue when the model exists.
 
Last edited:
Three said:
What does that even mean? A trained model is not really dependant on "DirectX and windows". It's running that trained model on different hardware/OS/int8 that requires work but you dismissed that time consuming work to mention 'training' as the issue when the model exists.
Click to expand...

LOL, you really think that a model like FSR4 just runs on thin air. :messenger_tears_of_joy:
 
winjer said:
LOL, you really think that a model like FSR4 just runs on thin air. :messenger_tears_of_joy:
Click to expand...
"running" is not the same as "training". FSR4.1's trained models aren't specific to PS5 Pro, windows Direct X, it didn't require separate training.

Running it on its hardware with only int8 RDNA2 is specific and that was exactly what you were dismissing as the reason for the time it took to come, in favour of 'training' being the reason you stated. Hence why people inferred you meant it required different "training" using int8. It's not hard to understand unless your skull runs on thin air.
 
Three said:
"running" is not the same as "training". FSR4.1's trained models aren't specific to PS5 Pro, windows Direct X, it didn't require separate training.

Running it on its hardware with only int8 RDNA2 is specific and that was exactly what you were dismissing as the reason for the time it took to come, in favour of 'training' being the reason you stated. Hence why people inferred you meant it required different "training" using int8. It's not hard to understand unless your skull runs on thin air.
Click to expand...

Even for training you need a complete software stack. From the OS, programs, programming languages, APIs, drivers, etc.
 
winjer said:
Even for training you need a complete software stack. From the OS, programs, programming languages, APIs, drivers, etc.
Click to expand...
Of course you do but you end up with a mathematical model that can be portable. Your model isn't strictly tied to a programming language, API, OS etc. You can make it work wherever/however you like. People have brought FSR4 to Vulkan, they didn't train new models for it did they?
 
Three said:
Of course you do but you end up with a mathematical model that can be portable. Your model isn't strictly tied to a programming language, API, OS etc. You can make it work wherever/however you like. People have brought FSR4 to Vulkan, they didn't train new models for it did they?
Click to expand...

Your model does nothing without connecting to the proper driver, API, OS.
 
Three said:
"running" is not the same as "training". FSR4.1's trained models aren't specific to PS5 Pro, windows Direct X, it didn't require separate training.

Running it on its hardware with only int8 RDNA2 is specific and that was exactly what you were dismissing as the reason for the time it took to come, in favour of 'training' being the reason you stated. Hence why people inferred you meant it required different "training" using int8. It's not hard to understand unless your skull runs on thin air.
Click to expand...
I'm confused, you're saying the pro is rdna2?
 
The PS5 Pro has a custom chip. The GPU is not available outside of the PS5 Pro to test. No direct equivalent that you can buy for PC. Neither is PSSR 2 available outside the PS5 Pro. Mark Cerny says it uses the same neural network as FSR 4.1 and he has no reason to lie. How it achieves this could be through specific hardware in the chip itself. It doesn't help that both AMD and Nvidia clearly hold back features to sell their latest GPU. It would be odd for the PS5 Pro to be capable of better upscaling than a 7800XTX which is why you get the disbelief around PSSR 2 being FSR 4.1 but that would be in keeping with with these companies.
 
"Great questions, particularly considering that FSR Frame Generation is technology that was co-developed between SIE and AMD, we're intimately familiar with it," Cerny revealed. "All I can say is that we have no more releases planned for this year. And that I look forward to discussing this more in the future!"
Click to expand...
This answers the question earlier why they took forever with PSSR2 winjer winjer .
Because they are slow.
 
I think it makes sense that AMD would use the same tech with Sony to make both FSR 4.1 and PSSR 2. I think we should just be happy that AMD cards and this the Pro now have great ML upscalers and, they're only gonna get better.
 
vkbest said:
www.digitalfoundry.net

Mark Cerny Confirms Frame Generation "Should Be Seen At Some Point On PlayStation Platforms"

"I look forward to discussing this more in the future!"
www.digitalfoundry.net www.digitalfoundry.net

This almost confirms no Frame Generation for Pro.
Click to expand...

I think this statement means it's coming next year in 2027.

"I'm very happy with how that work is progressing, and an equivalent frame generation library should be seen at some point on PlayStation platforms."
Click to expand...

It's hard for me to see them releasing it for PS6 and not PS5 Pro when most games in 2027 will be cross-gen.
 
mckmas8808 said:
I think this statement means it's coming next year in 2027.



It's hard for me to see them releasing it for PS6 and not PS5 Pro when most games in 2027 will be cross-gen.
Click to expand...
I think they would release this year specially for GTA VI if they are really doing on Pro
 
Nothing would stop devs from using FSR3 FG like Wukong does. It just wouldn't be the new Redstone ML FG. It's still quite good. Even AFMF is good when you are starting from higher FPS.
 
Buggy Loop said:
At that point, why would base PS5 not benefit, it's clearly not about TOPs that much.
Click to expand...

base PS5 is RDNA1 with RT hardware added, while Steam Deck is RDNA2... that's the main issue.

Xbox Series X however... that should work without much issue, and the increased number CUs there should help as well.
so it's kinda surprising that Microsoft isn't talking to AMD to bring FSR4 to Xbox consoles.
 
Last edited:
kevboard said:
base PS5 is RDNA1 with RT hardware added, while Steam Deck is RDNA2... that's the main issue.

Xbox Series X however... that should work without much issue, and the increased number CUs there should help as well.
so it's kinda surprising that Microsoft isn't talking to AMD to bring FSR4 to Xbox consoles.
Click to expand...

PS5's version of the GPU didn't include the "limited" but still available RDNA 2 ML ? I thought it did but those hardware debates from so many years ago I guess I forgot. They're fucked then if that's the case. What a missed opportunity.
 
Buggy Loop said:
PS5's version of the GPU didn't include the "limited" but still available RDNA 2 ML ? I thought it did but those hardware debates from so many years ago I guess I forgot. They're fucked then if that's the case. What a missed opportunity.
Click to expand...

afaik not even the PS5 Pro has all RDNA2 features. that's why PSSR uses a modified version of FSR4 that runs on their custom ML hardware.
base PS5 is RDNA1 at its core. the only important feature from RDNA2 that Sony wanted was the RT intersection accelerators
 
Last edited:
Buggy Loop said:
PS5's version of the GPU didn't include the "limited" but still available RDNA 2 ML ? I thought it did but those hardware debates from so many years ago I guess I forgot. They're fucked then if that's the case. What a missed opportunity.
Click to expand...

RDNA2 only has DP4A.
The PS5 doesn't even have that.
 
kevboard said:
afaik not even the PS5 Pro has all RDNA2 features. that's why PSSR uses a modified version of FSR4 that runs on their custom ML hardware.
Click to expand...

Then Microsoft is sleeping at the wheel if they have full RDNA 2 ML. It would be a nice upgrade for Series S and Series X. If you have it running on a Steam deck, there's no reasons those consoles would not benefit. Of course, might be a game to game basis depending on performances, but say, perhaps a game that used to be 60 fps would have FSR 4.1 in 40 fps mode (? realistic performance drop?) and I imagine a lot of peoples would be happy.
 
kevboard said:
Xbox Series X however... that should work without much issue, and the increased number CUs there should help as well.
so it's kinda surprising that Microsoft isn't talking to AMD to bring FSR4 to Xbox consoles.
Click to expand...

LOL

Series X GPU has 49 INT8 TOPS

How exactly could it run FSR4 when the minimum spec of RDNA4 is like 200???
 
Last edited:
Buggy Loop said:
Then Microsoft is sleeping at the wheel if they have full RDNA 2 ML. It would be a nice upgrade for Series S and Series X. If you have it running on a Steam deck, there's no reasons those consoles would not benefit. Of course, might be a game to game basis depending on performances, but say, perhaps a game that used to be 60 fps would have FSR 4.1 in 40 fps mode (? realistic performance drop?) and I imagine a lot of peoples would be happy.
Click to expand...

Microsoft had a program developing an ML upscaler, even before the release of this generation. They even showed a few pictures of it. But it never went anywhere.
Something like XeSS DP4A path can run on RDNA2, within 5-10% performance of FSR3.1, while looking significantly better. Had MS developed something similar, they could have a better upscaler than any other console, asides from the Pro.

FSR4 is significantly heavier. RDNA2 could use it for lower resolutions and smaller upscale factors.
But doing something like 1080 to full 4K in FSR4 could be troublesome.

This is an example of the FSR4 Int8 model leaked by AMD, running on a 6700XT. It's upscaling from 720p to 1080p.
But it has to use Optiscaler and a Vulkan translation layer, so there is some overhead. A proper optimized version could shave a few points of ms.
Be it on RDNA2 or RDNA3, this leaked dll of FSR4 int8, uses only DP4A. Though RNA3 has WMMA instructions to accelerate ML.

4zJB59zi9Anq8vKJ.jpg
 
Last edited:
Buggy Loop said:
Then Microsoft is sleeping at the wheel if they have full RDNA 2 ML. It would be a nice upgrade for Series S and Series X. If you have it running on a Steam deck, there's no reasons those consoles would not benefit. Of course, might be a game to game basis depending on performances, but say, perhaps a game that used to be 60 fps would have FSR 4.1 in 40 fps mode (? realistic performance drop?) and I imagine a lot of peoples would be happy.
Click to expand...

in most games they could drop the internal resolution while targeting the same output the game pushed for with FSR2/3 and be fine I bet.
the quality difference is absolutely insane, even on a low resolution system like the Steam Deck.

I posted comparisons in another thread a few weeks back

kevboard said:
just to show a quick comparison:

XeSS quality mode
pqkbtwr9.png



FSR4 balanced mode
z3h8ibp6.png


and in motion, the difference is even more in favor of FSR4.
both give me a solid 60fps (aside from UE4 typical I/O stutter)
Click to expand...

and that's XeSS vs FSR4, not even FSR3... FSR3 would look even worse than the XeSS example (the game doesn't support FSR2/3 tho... so...)
still shots also don't show how much better it looks in motion, where FSR4 completely shits on XeSS in this game.


XeSS Quality Mode:
pip4d33wMB0rJ7FI.jpg


FSR4 int8 Balanced Mode:
YS3v3LWJVB4G3Kn0.jpg
 
Last edited:
You must log in or register to reply here.

Similar threads

RDNA3 is getting FSR4.1 in July
2 3 4 5
247
1K
Gitaroo replied
AC Shadows PSSR2 TOGGLE vs PATCH
25
549
Bojji replied
[Digital Foundry] Mark Cerny confirms Frame Generation is coming to PlayStation at some point in the future
2 3
132
2K
Hawk The Slayer replied
[MLiD] PS6 & PSSR2 Dev Update
News Rumor Hype  4 5 6 7 8
388
6K
DeepEnigma replied
KeplerL2: PS6 Likely to Use 1TB SSD with No Disc Drive; Game Size Could Be Reduced with Neural Texture Compression
Opinion Rumor Hardware  2 3 4 5
200
2K
Mr.Phoenix replied
Top Bottom