Support NeoGAF

[viveks86]

This can't be system forced - titles that were tested against one version of PSSR can't be revalidated if system forces an update. However, I could see a rationale to allow users to select PSSR revision themselves - it's very PC and not at all console like, but then that ship sailed when the first Pro console launched, and we have HDR and other system settings now.

Yeah it can be a "use at your own risk" type of override in the future for those who want to tinker a bit (like most pro users). Can be presented similar to boost mode.

 

[Bojji]

Basically a port to different (custom) architecture of PS5 Pro. Will it have the same quality? We will see.

This also improves possibilities of FSR4 port to RDNA3 (or not, it doesn't support INT8).

 

[HeWhoWalks]

RIP PSSR I guess.
You will be missed.

"FSR 4 and this next evolution of PSSR are a paradigm for our future, going forward we expect to have our own implementations of each of the algorithms developed through the collaboration," says Cerny.

Only takes a moment of your time to read.

 

[Loomy]

I don’t understand why it takes until 2026 if fsr4 already exists ?

My guess is they're hoping to learn from the current FSR4 implementation. That will come with time. From there it's a matter of logistics. Where does it fit in with the work they're doing already, etc.

I'm guessing he also wants to launch it with more than a few games already making good use of it. Wouldn't be surprised if they want to do a better launch than the initial PSSR offering.

 

[vkbest]

I wonder if they will implement fake frames too in the next PSSR version in 2026

 

[S0ULZB0URNE]

That’s INT-4 as Bojji replied above. INT-8 without sparsity using very simple math is a quarter of that number.

RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RTX 3080: 238

INT-8 TOPs without sparsity.

Dat 4090 though.

 

[TheRedRiders]

TOPS are the new Teraflops.

LeviathanGamers2 take which I shared a few days ago seemed to have aged well, as Cerny confirmed that PS5 Pro and RDNA 4 have different ML architecture so directly porting FSR 4 to PS5 Pro wouldn't have been a straight forward process.

Shared this on the other thread too, and also worth sharing here. LeviathanGamer2's take on why Sony went with their own custom PSSR solution rather than plucking FSR 4.

He was also on the ball about Project Amethyst being undertaken to ensure that FSR4 and it's iterations would find their way onto PS5 Pro. Not anything prophetical I know but worth sharing.

 

[viveks86]

"FSR 4 and this next evolution of PSSR are a paradigm for our future, going forward we expect to have our own implementations of each of the algorithms developed through the collaboration," says Cerny.

Only takes a moment of your time to read.

I don't understand either side of this argument actually. It's like we are more invested in the brands and labels, instead of the underlying tech and results. Are we just trying to figure out who we should be praising or crediting for all the hard work? If the neural network and algorithms are common and just the target platform and implementation is different, it really is a case of "same.... but different... but same". Call it PSSR, FSR, PSFSR... whatever. How does it matter? It's born out a collaborative effort to achieve better results. I think people are spending too much time trying to mark this as a "win for Sony" or a "win for AMD", when it's actually a win for everybody that isn't Nvidia.

 

[HeWhoWalks]

I don't understand either side of this argument actually. It's like we are more invested in the brands and what it should be called, instead of the underlying tech and results. Are we just trying to figure out who we should be praising or crediting for all the hard work? If the neural network and algorithms are common and just the target platform and implementation is different, it really is a case of "same.... but different... but same". Call it PSSR, FSR, PSFSR... whatever. How does it matter? It's borne out a collaborative effort to achieve better results. I think people are spending too much time trying to mark this as a "win for Sony" or a "win for AMD", when it's actually a win for everybody that isn't Nvidia.

I don’t think you are aiming this post at the right user.

 

[bender]

It ONLY will take Pro 2 years to become good and worth buying.

 

[viveks86]

I don’t think you are aiming this post at the right user.

It wasn't aimed at you. I was just commenting on that debate as a whole (and I like your posts). Who cares if PSSR does or doesn't survive or if it should still be called PSSR or not, if the end result is a better upscaler where both Sony and AMD are involved? Sony will likely call it PSSR and AMD will call it FSR, but the underlying model will more or less be the same. This is a good thing, because from a game dev's standpoint, they really need to implement it once and can expect similar results on console and PC with AMD GPUs.

 

[HeWhoWalks]

It wasn't aimed at you. I was just commenting on that debate. Who cares if PSSR does or doesn't survive or if it should still be called PSSR or not, if the end result is a better upscaler where both Sony and AMD are involved? Sony will likely call it PSSR and AMD will call it FSR, but the underlying model will more or less be the same. This is a good thing, because from a game dev's standpoint, they really need to implement it once and can expect similar results on console and PC with AMD GPUs.

Oh, I totally agree. That’s why I don’t get statements such as “RIP PSSR”. I listened and read and was finding that completely at odds with what is happening!

And yes, NVIDIA needs this comp. I love DLSS, but it’s good to see the fight coming from the other direction too!

 

[ReBurn]

It will be interesting to see how close the PS5 Pro implementation gets to FSR4 on RDNA4 cards.

 

[Three]

Dat 4090 though.

The top end nvidia cards have high AI performance. The 5090 would be 838 if I'm not mistaking.

RTX 5090: 838
RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RX 9070: 290
RTX 3080: 238

The 9070 having lower performance and supporting FSR4 should show PS5 Pro is no slouch with the same workload though.

 

[Zathalus]

The top end nvidia cards have high AI performance. The 5090 would be 838 if I'm not mistaking.

RTX 5090: 838
RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RX 9070: 290
RTX 3080: 238

The 9070 having lower performance and supporting FSR4 should show PS5 Pro is no slouch with the same workload though.

Real world numbers are usually around 10-15% higher than those numbers, due to the way boost GPU clocks works on modern GPUs.

 

[Kataploom]

That sounds good but the issue is that many games will be stuck with the old PSSR. If they can update them all then great but if they dont then it's definitely disappointing. That's why I love pc, I dont depend on the developer to update the latest features, you can just do it yourself by easily modding games

Maybe PSSR implementations can be mapped to FSR4 automatically just like FSR 3.1 does on PC

 

[Radical_3d]

It's for Int4, so Int8 is half of that:

If this is with sparsity and Pro is not then base figures for for 9070XT is like ~390tops compared to 300 in Pro.

That's Int4.
And I am not 100% sure if that also uses sparsity - I'd first assumed no, but AMD played a bit coy by only giving AI multipliers to previous gen (4x normal, 8x with sparsity), and NVidia really muddied the waters with their PR here.

That’s INT-4 as Bojji replied above. INT-8 without sparsity using very simple math is a quarter of that number.

RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RTX 3080: 238

INT-8 TOPs without sparsity.

 

[viveks86]

Lol. But which part do you not get? 8 bit integer computation is twice as heavy as 4 bit computation. Using sparsity results in twice the effective compute. We need to ignore sparsity because PSSR does not use sparsity and it seems FSR4 does not need it either. So if you want to do an apples to apples comparison with the Pro, you should divide 1557 TOPs of 9070 XT by 4 (to account for 8 bit and sparsity), which is ~390, compared to the 300 of the Pro.

 

[Clear]

I don’t understand why it takes until 2026 if fsr4 already exists ?

Its because they essentially need to "port" the library to run on the Pro's APU. No joke because when neural nets require trillions of calculations to function everything has to be ultra-optimized. The traditional solution for this being to design the hardware component in such a way that this is able to be done flexibly and efficiently. With the PSU its seemingly a somewhat unique piece of hardware and may pose unique issues when it comes to implementation.

Remember the 300 tops is theoretical max throughput, so if applying the algorithm is essentialy like making a dog stand on its hind legs... then the actual throughput is likely to be lower, possibly to a point at which its too expensive to use.

 

[TGO]

And someone told me FSR4 wasn't compatible with PS5...

 

[Radical_3d]

Lol. But which part do you not get? 8 bit integer computation is twice as heavy as 4 bit computation. Using sparsity results in twice the effective compute. We need to ignore sparsity because PSSR does not use sparsity and it seems FSR4 does not need it either. So if you want to do an apples to apples comparison with the Pro, you should divide 1557 TOPs of 9070 XT by 4 (to account for 8 bit and sparsity), which is ~390, compared to the 300 of the Pro.

I don’t understand the concept of sparsity except when I look at my bank account.

 

[viveks86]

And someone told me FSR4 wasn't compatible with PS5...

It isn't. That's why they will take the rest of the year to make it so on the Pro. Will assume you are talking about the Pro and not the base PS5 as that has no ML hardware.

I don’t understand the concept of sparsity except when I look at my bank account.

Just think of it as an optimization technique that doubles performance. Like zipping a file and copying it increases how much you can effectively copy in a second, "sparse matrix computation" doubles neural net performance.

 

[Topher]

And someone told me FSR4 wasn't compatible with PS5...

I mean......based on the information we had before today FSR 4 was said to only work with RDNA 4.

 

[TGO]

I mean......based on the information we had before today FSR 4 was said to only work with RDNA 4.

True, but

Interestingly, the PS5 Pro GPU is still based on AMD's RDNA 2 (RX 6000-series) architecture, but with some features backported from RDNA 3 (RX 7000-series) - and from RDNA 4 (RX 8000-series, expected 2025) as well.

It's not exactly like for like though is it, and I guess it must have the features required for FSR4?

 

[Topher]

True, but

It's not exactly like for like though is it, and I guess it must have the features required for FSR4?

Ah....ok. The various architectures being incorporated is probably the reason for the confusion. Or....at least mine. lol

 

[viveks86]

It's not exactly like for like though is it, and I guess it must have the features required for FSR4?

FP8 computation specifically. We don't know if that was incorporated into the Pro's cross generational RDNA architecture or not, but regardless, it is not the same thing as a vanilla RDNA 4. So work needs to be done as it won't run out of the box.

 

[rickybooby87]

And someone told me FSR4 wasn't compatible with PS5...

It was me.

 

[Lokaum D+]

Did you bought a base PS5? Fuck you we have nothing for you anymore... Never buying base consoles ever again.

 

[James Sawyer Ford]

Its because they essentially need to "port" the library to run on the Pro's APU. No joke because when neural nets require trillions of calculations to function everything has to be ultra-optimized. The traditional solution for this being to design the hardware component in such a way that this is able to be done flexibly and efficiently. With the PSU its seemingly a somewhat unique piece of hardware and may pose unique issues when it comes to implementation.

Remember the 300 tops is theoretical max throughput, so if applying the algorithm is essentialy like making a dog stand on its hind legs... then the actual throughput is likely to be lower, possibly to a point at which its too expensive to use.

I guess I don’t understand why Sony didn’t put RDNA4 into PS5 Pro?

Is there something special about PS5 Pro’s hardware that has an advantage vs RDNA4?

I dont think RDNA4 would break BC, because surely PS6 will just use a new architecture and still have BC

 

[mckmas8808]

That’s INT-4 as Bojji replied above. INT-8 without sparsity using very simple math is a quarter of that number.

RTX 4090: 660
RTX 3090: 320
RX 9070 XT: 390
PS5 Pro: 300
RTX 3080: 238

INT-8 TOPs without sparsity.

And what about the OG PS5?

 

[TheRedRiders]

I guess I don’t understand why Sony didn’t put RDNA4 into PS5 Pro?

Is there something special about PS5 Pro’s hardware that has an advantage vs RDNA4?

I dont think RDNA4 would break BC, because surely PS6 will just use a new architecture and still have BC

PS5 Pro's ML hardware is a different implementation than what is found on RDNA 4, Sony likely went down their route because it meant that they could keep the silicon smaller and keep the costs down. RDNA 4 CU's are bigger than RDNA 3 CU's and in turn RDNA 2 CU's.

Due to the difference in ML hardware, it'll take a little more effort to implement FSR 4 onto an architecture it wasn't designed for.

 

[rofif]

nobody answered me before.
How is fsr4 different from pssr?
Isn't pssr Machine learning AI Upscaling just like dlss and fsr4 ?
I realize fsr2/3 were dumb solutions. But pssr finally transcended... so why we need fsr4 over it? if its the same?

 

[Gaiff]

I guess I don’t understand why Sony didn’t put RDNA4 into PS5 Pro?

Is there something special about PS5 Pro’s hardware that has an advantage vs RDNA4?

I dont think RDNA4 would break BC, because surely PS6 will just use a new architecture and still have BC

They did put RDNA4 for the ray tracing hardware. ML hardware is custom. GPU is RDNA-2 based. It's a mix and match of different components to maximize cost and efficiency. There's also time to consider. The PS5 Pro was released in November 2024 and the first RDNA4 GPU only hit the scene 4 months later. The chip would have also been physically larger.

And what about the OG PS5?

Don't know the number, but it has to be insanely low. The Series X only has 49 and I recall the PS5 having an even lower number or none at all.

nobody answered me before.
How is fsr4 different from pssr?
Isn't pssr Machine learning AI Upscaling just like dlss and fsr4 ?
I realize fsr2/3 were dumb solutions. But pssr finally transcended... so why we need fsr4 over it? if its the same?

Different neural network and training. FSR4 is more advanced and PSSR will eventually move to those newer and more advanced models to produce better results. As to how exactly? I haven't seen a direct comparison, but at a glance, FSR4 is significantly more stable and even better at retaining texture quality and fine details.

 

[Wolzard]

Some gaming news sites were saying that everyone should thank Sony for FSR 4.

Now the news comes out saying that Sony will adapt FSR 4 for the PS5 Pro.

 

[Wolzard]

They did put RDNA4 for the ray tracing hardware. ML hardware is custom. GPU is RDNA-2 based. It's a mix and match of different components to maximize cost and efficiency. There's also time to consider. The PS5 Pro was released in November 2024 and the first RDNA4 GPU only hit the scene 4 months later.

Don't know the number, but it has to be insanely low. The Series X only has 49 and I recall the PS5 having an even lower number or none at all.



Different neural network and training. FSR4 is more advanced and PSSR will eventually move to those newer and more advanced models to produce better results. As to how exactly? I haven't seen a direct comparison, but at a glance, FSR4 is significantly more stable and even better at retaining texture quality and fine details.

ML hardware is just WMMA from RDNA3.

How to accelerate AI applications on RDNA 3 using WMMA

This blog is a quick how-to guide for using the WMMA feature with our RDNA 3 GPU architecture using a Hello World example.

gpuopen.com

 

[Physiognomonics]

It was so smart by Cerny to wait for analysists to judge FSR4 before stating it was their API.

DF 5 days ago.

FSR4 is a A Big Leap Forward...PS5 pro has roughly 300 tops presumably with sparcity [4x less than FSR4 needs implying PS5 Pro would never be able to run it with no mention of FSR4 been possibly the first of Amethyst project.]

Cerny 5 hour ago. PS5 Pro 300 TOPs is without sparsity, Pro won't need sparsity for FSR4, which BTW we co-designed with AMD.

"The neural network (and training recipe) in FSR 4's upscaler are the first results of the Amethyst collaboration," Cerny told us. "And results are excellent, it's a more advanced approach that can exceed the crispness of PSSR. I’m very proud of the work of the joint team!"...We don't believe sparsity is useful for this particular upscaling algorithm.

Why are they always FUDing and lying about anything Playstation, always imagining the worst, never even once assuming FSR4 was co-designed by Sony etc. Just watch the tone of their articles about anything Xbox, always assuming the box would overperform when it almost always disappoint. I remember one video about them fantasizing like fanboys about a 2025 mid-gen Xbox that would magically performs like a 2027 console... They are MS propagandists!

 

[viveks86]

nobody answered me before.
How is fsr4 different from pssr?
Isn't pssr Machine learning AI Upscaling just like dlss and fsr4 ?
I realize fsr2/3 were dumb solutions. But pssr finally transcended... so why we need fsr4 over it? if its the same?

There are many types of AI upscaling. DLSS 2, 3 and PSSR use Convolutional Neural Networks (CNN). DLSS 4 uses Transformer neural networks. FSR 4 allegedly uses a combination of both. Since AMD and Sony are collaborating for the last year, Sony wants to take the underlying neural network(s) built for FSR 4 and incorporate them as the next evolution of PSSR. FSR4 and PSSR are currently not the same. After they hit the next big iteration, the core might be the same. They might be different cars with the same engine tuned for different use cases. Or same cars, tuned differently with just different labels. Or somewhere in between. At the minimum, they will be tuned for different desired results.

 

[SegaSnatcher]

FSR4 is AMD finally catching up to Nvidia. Nvidia still has an edge via its new Transformer model, but it looks like FSR4 still holds up even against direct comparisons with DLSS 4 having the edge in sharpness fine detail.

TLDR: PSSR 1.0 is yesterdays news, time for Sony to learn what they can from the all the advances seen in FSR4 and implement what they can.

 

[Danny Dudekisser]

God, I am so sick of this PSSR/any upscaling shit.

 

[Loxus]

I guess I don’t understand why Sony didn’t put RDNA4 into PS5 Pro?

Is there something special about PS5 Pro’s hardware that has an advantage vs RDNA4?

I dont think RDNA4 would break BC, because surely PS6 will just use a new architecture and still have BC

Might have something to do with sparsity.

PS5 Pro AI Accelerators may not have sparsity as Digital Foundry stated the 300TOPs is without sparsity.

The 9070 has 583 INT8 TOPs with sparsity, so the PS5 Pro 300TOPs may not have been enough to run FSR4 and needed more time to optimize PSSR to get near FSR4 with what they have.

RDNA3 has 1st gen AI Accelerators while RDNA4 has 2nd gen, PS5 Pro may be in between them both resulting in it not having sparsity.

Just my speculation.

 

[James Sawyer Ford]

Might have something to do with sparsity.

PS5 Pro AI Accelerators may not have sparsity as Digital Foundry stated the 300TOPs is without sparsity.

The 9070 has 583 INT8 TOPs with sparsity, so the PS5 Pro 300TOPs may not have been enough to run FSR4 and needed more time to optimize PSSR to get near FSR4 with what they have.

RDNA3 has 1st gen AI Accelerators while RDNA4 has 2nd gen, PS5 Pro may be in between them both resulting in it not having sparsity.

Just my speculation.

This was shot down

 

[rofif]

They did put RDNA4 for the ray tracing hardware. ML hardware is custom. GPU is RDNA-2 based. It's a mix and match of different components to maximize cost and efficiency. There's also time to consider. The PS5 Pro was released in November 2024 and the first RDNA4 GPU only hit the scene 4 months later. The chip would have also been physically larger.

Don't know the number, but it has to be insanely low. The Series X only has 49 and I recall the PS5 having an even lower number or none at all.



Different neural network and training. FSR4 is more advanced and PSSR will eventually move to those newer and more advanced models to produce better results. As to how exactly? I haven't seen a direct comparison, but at a glance, FSR4 is significantly more stable and even better at retaining texture quality and fine details.

and silent hill 2 forever will stay on broken pssr version.... I just know it

 

[viveks86]

And what about the OG PS5?

I wouldn't even count it as it's of no use for upscaling. It's from WMMA which helps accelerate shaders or perform physics simulations. Also helps accelerate AI applications that aren't graphics driven. It really wasn't meant for the bandwidth/computations needed for AI upscaling. Several architectural changes had to be made to enable it on the Pro. Cerny goes pretty in depth on it on the previous interview with DF and how he would like to improve on it for the PS6. For all intents and purposes, we should assume the base consoles can't do any realtime AI-driven upscaling.

 

[James Sawyer Ford]

It was so smart by Cerny to wait for analysists to judge FSR4 before stating it was their API.

DF 5 days ago.


Cerny 5 hour ago. PS5 Pro 300 TOPs is without sparsity, Pro won't need sparsity for FSR4, which BTW we co-designed with AMD.


Why are they always FUDing and lying about anything Playstation, always imagining the worst, never even once assuming FSR4 was co-designed by Sony etc. Just watch the tone of their articles about anything Xbox, always assuming the box would overperform when it almost always disappoint. I remember one video about them fantasizing like fanboys about a 2025 mid-gen Xbox that would magically performs like a 2027 console... They are MS propagandists!

What’s also a bit annoying is Oliver’s tone on current PSSR

It’s so absurdly negative, claiming it’s not much better than fsr2

And it’s pretty far from the reality. It’s not as good as other techniques but it’s still leagues better than other solutions and transformational for many titles like stellar blade, ff7r, and tlou

 

[jroc74]

I just posted this response the other day..

 

[Lysandros]

What’s also a bit annoying is Oliver’s tone on current PSSR

It’s so absurdly negative, claiming it’s not much better than fsr2

And it’s pretty far from the reality. It’s not as good as other techniques but it’s still leagues better than other solutions and transformational for many titles like stellar blade, ff7r, and tlou

"Competitive with FSR 2 at least", what a farce. And they gave him the honor to interview Marc Cerny.

 

[Bojji]

What’s also a bit annoying is Oliver’s tone on current PSSR

It’s so absurdly negative, claiming it’s not much better than fsr2

And it’s pretty far from the reality. It’s not as good as other techniques but it’s still leagues better than other solutions and transformational for many titles like stellar blade, ff7r, and tlou

Both of those don't have any reconstruction on base PS5 so they would look much better on 4k screen with FSR2 vs. what they have. Of course PSSR is better than FSR2 but it's still the weakest ML upscaler:

So news about achieving FSR4 quality is very good. Question is, will it have the same performance impact as current PSSR?

 

[vkbest]

Lol. But which part do you not get? 8 bit integer computation is twice as heavy as 4 bit computation. Using sparsity results in twice the effective compute. We need to ignore sparsity because PSSR does not use sparsity and it seems FSR4 does not need it either. So if you want to do an apples to apples comparison with the Pro, you should divide 1557 TOPs of 9070 XT by 4 (to account for 8 bit and sparsity), which is ~390, compared to the 300 of the Pro.

Sparsity in the best scenario can improve a 20% to 30%. Also FSSR4 is not using Int4. Also 9070 not XT is around 290-291 TOPS.

 

[mansoor1980]

"Competitive with FSR 2 at least", what a farce. And they gave him the honor to interview Marc Cerny.

What’s also a bit annoying is Oliver’s tone on current PSSR

It’s so absurdly negative, claiming it’s not much better than fsr2

And it’s pretty far from the reality. It’s not as good as other techniques but it’s still leagues better than other solutions and transformational for many titles like stellar blade, ff7r, and tlou

 

[James Sawyer Ford]