Support NeoGAF

[Karim]

Because in your mind AMD developing FSR2 helped them in any way with FSR4. This isn't true at all, those things are separate just like Insomniac reconstruction is to PSSR.

Listen you are right. Whatever.

 

[Salty Pickle]

Question is will dual issue compute offer a meaningful difference in such workloads considering it's bandwidth heavy, and the PS6 bandwidth isn't as impressive as its other specs, I think less than 50% improvement over the Pro?

But isn't rdna5 supposed to be way more efficient than rdna4 with bandwidth?

 

[Shmunter]

Because in your mind AMD developing FSR2 helped them in any way with FSR4. This isn't true at all, those things are separate just like Insomniac reconstruction is to PSSR.

Correct, FA4 is the first ML upscaler. Completely different approach to the special and temporal ones.



FSR 1 (2021): Spatial upscaling (algorithmic, single-frame only).
Analyzes edges and shapes in the current low-res frame to reconstruct a higher-res image. No use of previous frames (non-temporal), no motion vectors, no AI/ML. Lowest image quality at aggressive settings compared to later versions. No frame generation.


• FSR 2 (2022): Temporal upscaling (algorithmic).
Uses frame history (previous frames + motion vectors + depth data) for accumulation and reconstruction. Much better quality and anti-aliasing than FSR 1. Still fully non-AI (shader-based). No frame generation (though some games later added it via other means). Superseded by FSR 3/3.1.


• FSR 3 (2023): Temporal upscaling (improved version of FSR 2's algorithm) + Frame Generation.
Core upscaler remains algorithmic/temporal (no AI). Frame generation uses optical flow (from AMD Fluid Motion Frames) to interpolate entirely new frames analytically. Big FPS boost but tied to FSR's own upscaler at launch. Still shader-only (no ML).


• FSR 3.1 (2024): Enhanced temporal upscaling (algorithmic, with major quality tweaks) + decoupled Frame Generation.
Same non-AI temporal foundation as FSR 3/2 but with big improvements in detail preservation, reduced ghosting, better temporal stability, and less flickering/shimmering. Frame generation is now fully separate and works with any upscaler (DLSS, XeSS, native TAA, etc.). Still fully algorithmic—no AI involved.


• FSR 4 / Redstone (2025–2026, now officially "AMD FSR Upscaling" in the Redstone suite): AI/ML-powered upscaling (machine-learning accelerated, temporal + neural network) + ML-powered Frame Generation.
First version to use AI/ML (trained neural networks for reconstruction). Delivers significantly sharper results and fewer artifacts than FSR 3.1. Exclusive to RDNA 4 GPUs (Radeon RX 9000 series) for the full ML features; older GPUs fall back to the analytical FSR 3.1 version. Includes extra ML tools like Ray Regeneration. This is the current flagship.

 

[Karim]

Correct, FA4 is the first ML upscaler. Completely different approach to the special and temporal ones.



FSR 1 (2021): Spatial upscaling (algorithmic, single-frame only).
Analyzes edges and shapes in the current low-res frame to reconstruct a higher-res image. No use of previous frames (non-temporal), no motion vectors, no AI/ML. Lowest image quality at aggressive settings compared to later versions. No frame generation.


• FSR 2 (2022): Temporal upscaling (algorithmic).
Uses frame history (previous frames + motion vectors + depth data) for accumulation and reconstruction. Much better quality and anti-aliasing than FSR 1. Still fully non-AI (shader-based). No frame generation (though some games later added it via other means). Superseded by FSR 3/3.1.


• FSR 3 (2023): Temporal upscaling (improved version of FSR 2's algorithm) + Frame Generation.
Core upscaler remains algorithmic/temporal (no AI). Frame generation uses optical flow (from AMD Fluid Motion Frames) to interpolate entirely new frames analytically. Big FPS boost but tied to FSR's own upscaler at launch. Still shader-only (no ML).


• FSR 3.1 (2024): Enhanced temporal upscaling (algorithmic, with major quality tweaks) + decoupled Frame Generation.
Same non-AI temporal foundation as FSR 3/2 but with big improvements in detail preservation, reduced ghosting, better temporal stability, and less flickering/shimmering. Frame generation is now fully separate and works with any upscaler (DLSS, XeSS, native TAA, etc.). Still fully algorithmic—no AI involved.


• FSR 4 / Redstone (2025–2026, now officially "AMD FSR Upscaling" in the Redstone suite): AI/ML-powered upscaling (machine-learning accelerated, temporal + neural network) + ML-powered Frame Generation.
First version to use AI/ML (trained neural networks for reconstruction). Delivers significantly sharper results and fewer artifacts than FSR 3.1. Exclusive to RDNA 4 GPUs (Radeon RX 9000 series) for the full ML features; older GPUs fall back to the analytical FSR 3.1 version. Includes extra ML tools like Ray Regeneration. This is the current flagship.

Different approach doesn't mean you haven't learned better logic to improve the upscaler algorithmic job.

 

[Felessan]

And where is the PSSR equivalent of FSR4.0? There isn't because PSSR and FSR4 were designed separately.

There isn't because FSR4 not possible on Int8. It's not possible. Due to internal structure of basis transformer that explicitly requires fp calculations

Show me any vendor transformer that run on Int8, released in 2024 - there is none, it was not possible then, research of models didn't reach that stage yet.

Mumbling about "easy to move from fp8 to Int8" shows that you know literally nothing about this stuff

FSR4.1 is the first result of Amethyst.

No it isnt. See Cerny interview above.

There is no version of FSR4 that looks as bad as PSSR1. And there is no version of PSSR1 that looks as good as FSR4.0
But there is now a version of PSSR2 that looks as good as FSR4.1.

There is no version of FSR4 because AMD opted to skip use CNN model, by the time they have ML capable hardware fp8 transformer already was finished

 

[Bojji]

There isn't because FSR4 not possible on Int8. It's not possible. Due to internal structure of basis transformer that explicitly requires fp calculations

Show me any vendor transformer that run on Int8, released in 2024 - there is none, it was not possible then, research of models didn't reach that stage yet.

Mumbling about "easy to move from fp8 to Int8" shows that you know literally nothing about this stuff


No it isnt. See Cerny interview above.


There is no version of FSR4 because AMD opted to skip use CNN model, by the time they have ML capable hardware fp8 transformer already was finished

So how int8 version of FSR4 exists?

 

[winjer]

There isn't because FSR4 not possible on Int8. It's not possible. Due to internal structure of basis transformer that explicitly requires fp calculations

Show me any vendor transformer that run on Int8, released in 2024 - there is none, it was not possible then, research of models didn't reach that stage yet.

Mumbling about "easy to move from fp8 to Int8" shows that you know literally nothing about this stuff


No it isnt. See Cerny interview above.


There is no version of FSR4 because AMD opted to skip use CNN model, by the time they have ML capable hardware fp8 transformer already was finished

FSR4 is not possible on the Pro because it was made for RDNA4, by AMD.
That is why there was never an FSR4.0 version on the Pro.

And FSR still uses mostly CNN model. They only use Transformer operations in a limited scale.
And a Transformer model is not exclusive to FP8. DLSS4 is a full Transformer model and it uses both Int8 and FP8 operations.

 

[Felessan]

So how int8 version of FSR4 exists?

There is no int8 FSR4
PSSR2 is the same model as FSR4.1 which is almost an year later than FSR4

 

[Bojji]

There is no int8 FSR4
PSSR2 is the same model as FSR4.1 which is almost an year later than FSR4

So what is this?

 

[winjer]

There is no int8 FSR4
PSSR2 is the same model as FSR4.1 which is almost an year later than FSR4

LOL, the leaked FSR4 model was Int8. That is how it can run in RDNA2 and RDNA4 in DP4A.

 

[Felessan]

FSR4 is not possible on the Pro because it was made for RDNA4, by AMD.
That is why there was never an FSR4.0 version on the Pro.

You keep repeating bullshit
I am not talking about FSR4, I am talking of transformer circa 2024 version - find me a proof that any, not necessary Pro, int8 hardware can run that model

And FSR still uses mostly CNN model. They only use Transformer operations in a limited scale.
And a Transformer model is not exclusive to FP8. DLSS4 is a full Transformer model and it uses both Int8 and FP8 operations.

They still use fp8 to run parts of model
In model circa 2024 you may as well abandon transformer altogether and go back to CNN if you can't run fp8, reverting back to pssr1 quality

 

[Felessan]

LOL, the leaked FSR4 model was Int8. That is how it can run in RDNA2 and RDNA4 in DP4A.

It most likely same 4.1 model PSSR2 use
It's still under "FSR4" first digit umbrella, but the model itself is newer

 

[winjer]

You keep repeating bullshit
I am not talking about FSR4, I am talking of transformer circa 2024 version - find me a proof that any, not necessary Pro, int8 hardware can run that model

DLSS4 uses int8 and FP8 instructions. And it's a Transformer model.
Only DLSS4.5 uses only FP8.

They still use fp8 to run parts of model
In model circa 2024 you may as well abandon transformer altogether and go back to CNN if you can't run fp8, reverting back to pssr1 quality

Because it was made to run on RDNA4.

 

[vkbest]

It most likely same 4.1 model PSSR2 use
It's still under "FSR4" first digit umbrella, but the model itself is newer

You literally don't know what PSSR2 is doing. Could be even a FP16 model like the Sony's patent

 

[Moridianae]

I'm just very happy to see what has happened with the release, honestly. Up until this point, I'd seen no reason to justify a Pro(for myself).

But seeing this has me reconsidering it, especially considering I recall reading from a developer recently(it may have been on purple site, alas) that PSSR2 is not the most recent version, or something to that effect.

I'm assuming we'll see further improvements and refinement this year on it.

If I can trade in my old PS5 and snag a deal on a pre-owned Pro, might be worth upgrading.

 

[Bojji]

I'm just very happy to see what has happened with the release, honestly. Up until this point, I'd seen no reason to justify a Pro(for myself).

But seeing this has me reconsidering it, especially considering I recall reading from a developer recently(it may have been on purple site, alas) that PSSR2 is not the most recent version, or something to that effect.

I'm assuming we'll see further improvements and refinement this year on it.

If I can trade in my old PS5 and snag a deal on a pre-owned Pro, might be worth upgrading.

In some games having good ML reconstruction vs. no image reconstruction is like a nigh/day difference. Crimson Desert, RE9 and FFVII Rebirth look like shit on base PS5.

 

[Mr Moose]

Amethyst collab started in 2023 for PSSR and FSR4.

 

[Three]

And where is the PSSR equivalent of FSR4.0? There isn't because PSSR and FSR4 were designed separately.
FSR4.1 is the first result of Amethyst.
There is no version of FSR4 that looks as bad as PSSR1. And there is no version of PSSR1 that looks as good as FSR4.0
But there is now a version of PSSR2 that looks as good as FSR4.1.

This is like saying where is the FSR3 equivalent of FSR4.

The PSSR equivalent of FSR4 was PSSR2/FSR4.1. They were using the same collaborative development and training along the way though. It's just that when they set their release date for the major update to PS (using int8) it coincided with 4.1 and not 4.0.

Would you say there was no collaboration between AMD/PS on PS5/RDNA2 because there is no PlayStation equivalent of RDNA3 but RDNA5/PS6 exists?

 

[winjer]

This is like saying where is the FSR3 equivalent of FSR4.

The PSSR equivalent of FSR4 was PSSR2/FSR4.1. They were using the same collaborative development and training along the way though. It's just that when they set their release date for the major update to PS it coincided with 4.1 and not 4.0.

Would you say there was no collaboration between AMD/PS on PS5/RDNA2 because there is no PlayStation equivalent of RDNA3 but RDNA5/PS6 exists?

I would say that Sony and AMD were developing their own upscalers separately. Then realized they could save time and money by working together. Espeically Sony, which had the made the worst ML upscaler.
So they joined forces at the end of 2025, announced Project Amethyst and now we are seeing the first result: FSR4.1

BTW, did you know that AMD and Sony never collaborated all that much developing software?
For example, did you know that AMD does not provide drivers for Sony's consoles? And of course, Sony does not provide drivers for AMD.

 

[Moridianae]

I'm interested in knowing just how far they can advance PSSR on the Pro.

I'm excited to see what it can do on the PS6, for sure also.

 

[Felessan]

DLSS4 uses int8 and FP8 instructions. And it's a Transformer model.
Only DLSS4.5 uses only FP8.

If it use both, it requires both

Because it was made to run on RDNA4.

As if you can always write model to hardware
Why do hardware level optimization aka tailoring hardware to software paths if it's easy to write model for given hardware

 

[Three]

I would say that Sony and AMD were developing their own upscalers separately.

based on what when both AMD and Mark Cerny have said they worked on FSR4 jointly?

Then realized they could save time and money by working together. Espeically Sony, which had the made the worst ML upscaler.
So they joined forces at the end of 2025,

How the hell are you jumping to conclusions and ignoring the fact that Project Amethyst was publicly announced in 2024 and likely worked on even earlier?

BTW, did you know that AMD and Sony never collaborated all that much developing software?
For example, did you know that AMD does not provide drivers for Sony's consoles? And of course, Sony does not provide drivers for AMD.

Sure I guess Sony have to write their FreeBSD based drivers but what's the relevance?

 

[Mr Moose]

I would say that Sony and AMD were developing their own upscalers separately. Then realized they could save time and money by working together. Espeically Sony, which had the made the worst ML upscaler.
So they joined forces at the end of 2025, announced Project Amethyst and now we are seeing the first result: FSR4.1

BTW, did you know that AMD and Sony never collaborated all that much developing software?
For example, did you know that AMD does not provide drivers for Sony's consoles? And of course, Sony does not provide drivers for AMD.

This is false. They started in 2023.

"There are really two goals, one shorter term and one longer term," explains Cerny. "The shorter-term goal is to co-develop neural network architectures and training strategies for game graphics. We're stronger together than apart, so it makes a lot of sense to combine some of our resources when tackling these problems. And because we began this collaboration in earnest in late 2023 (when PSSR development was wrapping up), I'm happy to say that there have already been results.

FSR4.0 is part of that collab.

 

[winjer]

based on what when both AMD and Mark Cerny have said they worked on FSR4 jointly?

How the hell are you jumping to conclusions and ignoring the fact that Project Amethyst was publicly announced in 2024 and likely worked on even earlier?

Sure I guess Sony have to write their FreeBSD based drivers but what's the relevance?

If that is true, then where is the equivalent of FSR4.0 on the PSSR. There isn't.
PSSR1 and FSR4 are very different upscalers.

 

[Mr Moose]

If that is true, then where is the equivalent of FSR4.0 on the PSSR. There isn't.
PSSR1 and FSR4 are very different upscalers.

I don't think PSSR1 is part of Amethyst (or at least not fully and not the earliest version of PSSR1), but FSR 4.0 is.

 

[winjer]

I don't think PSSR1 is part of Amethyst (or at least not fully and not the earliest version of PSSR1), but FSR 4.0 is.

Then where is the FSR4 equivalent? Why is there only FSR4.1 and PSSR2 equivalent?

 

[Three]

If that is true, then where is the equivalent of FSR4.0 on the PSSR. There isn't.
PSSR1 and FSR4 are very different upscalers.

You're really arguing in circles here. Why would there be a FSR4.0 version when FSR4.1 released on PS first with PSSR2? FSR4 didn't come because when they were developing the PS5 Pro int8 algorithms they coincided with FSR4.1. For the same exact reason PS6 is RDNA5 and there is no PS equivalent of RDNA3. Doesn't mean PS5 wasn't jointly developed with AMD and RDNA2.

You even said yourself the training on FSR4 was done right? so what exactly was the hold up in your eyes other than having to write for PS5 Pro (int8)? Why didn't they just use FSR4?

 

[winjer]

You're really arguing in circles here. Why would there be a FSR4.0 version when FSR4.1 released on PS first with PSSR2? FSR4 didn't come because when they were developing the PS5 Pro int8 algorithms they coincided with FSR4.1. For the same exact reason PS6 is RDNA5 and there is no PS equivalent of RDNA3. Doesn't mean PS5 wasn't jointly developed with AMD and RDNA2.

FSR4.0 was released a year ago. And for a whole year, there was not a PSSR equivalent.
But there is one now: FSR 4.1

 

[Mr Moose]

Then where is the FSR4 equivalent? Why is there only FSR4.1 and PSSR2 equivalent?

No clue, but we know they started fully with it in 2023, and he said when PSSR development was wrapping up so that seems to suggest PSSR1 was not part of Amethyst.

And because we began this collaboration in earnest in late 2023 (when PSSR development was wrapping up), I'm happy to say that there have already been results.

From 2023 to Pro release, and then until now with PSSR2, they took their sweet ass time with it (PSSR1).

 

[winjer]

No clue, but we know they started fully with it in 2023, and he said when PSSR development was wrapping up so that seems to suggest PSSR1 was not part of Amethyst.
From 2023 to Pro release, and then until now with PSSR2, they took their sweet ass time with it (PSSR1).

Developing and training a model takes years. AMD was already full on developing FSR4 for RDNA4.
So Sony entering would only pay dividens years later, as in 2026 with FSR 4.1
Do you really think Sony would not want to have an upscaler of the quality of FSR4 a year ago? Of course they would, but it wasn't ready for the Pro. It was only ready for RDNA4.
And it took another year for Amethyst to have some result and that is FSR4.1

 

[Three]

FSR4.0 was released a year ago. And for a whole year, there was not a PSSR equivalent.
But there is one now: FSR 4.1

Yes, because they had to write those int8 algorithms for PS5 Pro specifically and implement/test them in games.

So explain it then, you're saying the training was done and that "converting to int8" isn't time consuming. What is the reason you think that they didn't just have FSR4 and went with FSR4.1 instead?

It was clearly their release schedule for PS support much like their hardware release schedule doesn't mean they didn't collaborate on RDNA just because PS6 is RDNA5 and there is no RDNA4 PlayStation.

 

[winjer]

Yes, because they had to write those int8 algorithms for PS5 Pro specifically and implement/test them in games.

So explain it then, you're saying the training was done and that "converting to int8" isn't time consuming. What is the reason you think that they didn't just have FSR4 and went with FSR4.1 instead?

It was clearly their release schedule for PS support much like their hardware release schedule doesn't mean they didn't collaborate on RDNA just because PS6 is RDNA5 and there is no RDNA4 PlayStation.

So Sony was developing FSR4 only for RDNA4, this whole time?

 

[Nex240]

it's always been obvious that PSSR is more or less a Github fork of FSR4 optimised for INT8. Sony just calls it PSSR for marketing.
Im surprised Nintendo didn't try to pretend DLSS was their own tech and call it NSSR or something, but Nvidia might have told them a big fat No.

 

[Mr Moose]

Developing and training a model takes years. AMD was already full on developing FSR4 for RDNA4.
So Sony entering would only pay dividens years later, as in 2026 with FSR 4.1
Do you really think Sony would not want to have an upscaler of the quality of FSR4 a year ago? Of course they would, but it wasn't ready for the Pro. It was only ready for RDNA4.
And it took another year for Amethyst to have some result and that is FSR4.1

4.0 is part of the collab. Why they took so long to put it to use (4.1) on the Pro I have no idea.

it's always been obvious that PSSR is more or less a Github fork of FSR4 optimised for INT8. Sony just calls it PSSR for marketing.
Im surprised Nintendo didn't try to pretend DLSS was their own tech and call it NSSR or something, but Nvidia might have told them a big fat No.

Not sure about PSSR1 from Cerny's comment, or at least the earliest version of it.

 

[Three]

So Sony was developing FSR4 only for RDNA4, this whole time?

No because they codeveloped it for PSSR2 too. I don't understand what your point of contention is.

 

[winjer]

No because they codeveloped it for PSSR2 too. I don't understand what your point of contention is.

Then where is FSR4.0 on the Pro?

 

[DeepEnigma]

it's always been obvious that PSSR is more or less a Github fork of FSR4 optimised for INT8. Sony just calls it PSSR for marketing.
Im surprised Nintendo didn't try to pretend DLSS was their own tech and call it NSSR or something, but Nvidia might have told them a big fat No.

They codeveloped it, they can call it whatever they want. Nintendo had zero co development in DLSS.

 

[Physiognomonics]

4.0 is part of the collab. Why they took so long to put it to use (4.1) on the Pro I have no idea.

Well they probably wanted to use 4.1 from the start on Pro and not waste their time with 4.0. Maybe they had to wait a little before 4.1 was finished. It's pure speculation.

 

[Nex240]

They codeveloped it, they can call it whatever they want. Nintendo had zero co development in DLSS.

Yes but Switch 2 DLSS is using it's own profile. If Nvidia was a smaller company like in the early 2000's Nintendo would have probably tried to market it as their own tech.

 

[Three]

Then where is FSR4.0 on the Pro?

Why would there be FSR4.0 on the Pro when FSR4.1 is the PSSR2 release? (coinciding with when they were ready with int8 and after Redstone enhancements)

You have both AMD and Mark telling you they collaborated on FSR4 yet you're still here arguing and asking this odd question. not even coming up with a possible reason yourself yet dismissing the idea that it had to be implemented on the PS5 pro int8 hardware.

 

[DeepEnigma]

Yes but Switch 2 DLSS is using it's own profile. If Nvidia was a smaller company like in the early 2000's Nintendo would have probably tried to market it as their own tech.

You got it all figured out.

 

[winjer]

Why would there be FSR4.0 on the pro when FSR4.1 is the PSSR2 release? (coinciding with when they were ready with int8 and after Redstone enhancements)

Because then the Pro could have a good upscaler in early 2025, instead of the lackluster PSSR1.

You have both AMD and Mark telling you they collaborated on FSR4 yet you're still here asking this odd question and not even coming up with a reason yourself yet dismissing the idea that it had to be implemented on the PS5 pro int8 hardware.

We also had AMD saying that the PS5 was RDNA2, when in reality it's just RDNA1 with RT units.
Marketing people have a way with words and reality...

 

[Three]

Because then the Pro could have a good upscaler in early 2025, instead of the lackluster PSSR1.

Ok but what has this got to do with what you were claiming? You were claiming FSR4 was separate from Amethyst. That Amethyst came after FSR4 when it didn't.

We also had AMD saying that the PS5 was RDNA2, when in reality it's just RDNA1 with RT units.
Marketing people have a way with words and reality...

Not this shit again. Both XSX and PS5 are custom RDNA. None of them have the full RDNA 2 features but they have the important features of RDNA2 (you know the time of hardware finalisation much like FSR support). Does that mean there was no collab and that RDNA2 and PS5's chip development are completely separate things? No.

 

[winjer]

Ok but what has this got to do with what you were claiming? You were claiming FSR4 was separate from Amethyst. That Amethyst came after FSR4 when it didn't.

Not this shit again. Both XSX and PS5 are custom RDNA. None of them have the full RDNA 2 features but they have the important features of RDNA2 (you know the time of hardware finalisation much like FSR support). Does that mean there was no collab and that RDNA2 and PS5's chip development are completely separate things? No.

The Series S/X have all the RDNA2 feature set. It's only missing the L3 cache.
The PS5 really is RDNA1 with RT units.

 

[Three]

The Series S/X have all the RDNA2 feature set. It's only missing the L3 cache.
The PS5 really is RDNA1 with RT units.

You're clearly just arguing for arguments sake.

 

[winjer]

You're clearlt just arguing for arguments sake.

Wait, you really think the PS5 is RDNA2? Seriously?

 

[Three]

Wait, you really think the PS5 is RDNA2? Seriously?

As I said they're custom RDNA closer to RDNA2 than any other version. Ultimately it's a pointless argument in practical terms and has nothing to do with the fact you're now dodging regarding FSR4 and the Sony/AMD collaboration Amethyst.

 

[winjer]

As I said they're custom RDNA closer to RDNA2 than any other version. Ultimately it's a pointless argument in practical terms and has nothing to do with the fact you're now dodging regarding FSR4 and the Sony/AMD collaboration Amethyst.

It's custom, but it's closer to RDNA1, than to RDNAA2.
It lacks all the RDNA2 features. The RT units are IP1.0, while RDNA2 is IP1.1
And even the ROPs are not RDNA2, they are RDNA1.
For someone who talks so much about tech, one would expect you would know a lot more.
And it's pointless to continuing arguing about FSR4/PSSR1. You have stated your mind and so have I. We are now just repeating the same arguments.

 

[Mr.Phoenix]

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.

You can't train an Ai model in Int8. All training is done in FP.

 

[winjer]

You can't train an Ai model in Int8. All training is done in FP.

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.