Support NeoGAF

[Karim]

Why can't you admit that Sony made a mistake by not having ML capabilities on the PS5? Even the most basic solution, DP4A, would be e net positive for the console and was readily available at the time.

Why you continue to change the argument? It's just that your real point? Prove ps5 is bad designed? Looking at the series X I think Sony has been smarter to choose such way and the market seem give them the reason. The ML capabilities on Series X are a waste of tech but this is just my opinion.

 

[winjer]

Why you continue to change the argument? It's just that your real point? Prove ps5 was bad designed?

I just said, it doesn't have support for DP4A, which would allow it to run a much better upscaler than FSR3.

 

[Poppyseed]

Saros does look lovely on the Pro. Definitely not always locked 60 though.

 

[Karim]

I just said, it doesn't have support for DP4A, which would allow it to run a much better upscaler than FSR3.

What upscaler could handle ps5 with such tech? FSR4? It take 11mss on Series X gpu theorically.

 

[MisterXDTV]

This thread aged like milk

 

[winjer]

What upscaler could handle ps5 with such tech? FSR4?

AMD already said it's going to run on RDNA2. For example.

 

[Karim]

AMD already said it's going to run on RDNA2. For example.

Doesn't means an hardware like Series X, for example, could run it at acceptable mss costs in the gpu. Just sayin. On console even AF filter still has absurd costs compared the pc.

 

[Bojji]

What upscaler could handle ps5 with such tech? FSR4? It take 11mss on Series X gpu theorically.

XeSS is better than FSR3 and runs decently on RDNA2 and 3. So something like this could always be done, shame on MS for dropping their ML upscaler mid development.

This thread aged like milk

Price also went up.

 

[Karim]

XeSS is better than FSR3 and runs decently on RDNA2 and 3. So something like this could always be done, shame for MS dropping their ML upscaler mid development.

You people continue to use something very hypothetical as it is very easy. Ah if MS ah if sony. Almost surely MS abandoned it because was a waste of time.

 

[winjer]

Doesn't means on Series X could run it at acceptable mss costs in the gpu. Just sayin.

It can. People are already running it with the FSR 4.0.2C with great success on DNA2 GPUs.
And this is not even an optimized dll, that runs directly from the driver without Optiscaler overhead.
Also, consider that XeSS DP4A is a distilled version of XeSS XMX, made to run less intensively on GPUs without dedicated ML units.
Sony and/or AMD could do it. In fact, this might just be what AMD is preparing for FSR4 on RDNA2. A distilled version that runs a little better, sacrificing a bit of quality.
Again, it would look miles better than FSR3.

 

[Bojji]

You people continue to use something very hypothetical as it is very easy.

XeSS is not hypothetical, same with FSR4.1. Both are real things that run on RDNA2 hardware (will run in the case of FSR - officially).

 

[Plies]

We'll get the FINAL verdict once GTA 6 comes out.

Then and only then..

 

[Karim]

It can. People are already running it with the FSR 4.0.2C with great success on DNA2 GPUs.
And this is not even an optimized dll, that runs directly from the driver without Optiscaler overhead.
Also, consider that XeSS DP4A is a distilled version of XeSS XMX, made to run less intensively on GPUs without dedicated ML units.
Sony and/or AMD could do it. In fact, this might just be what AMD is preparing for FSR4 on RDNA2. A distilled version that runs a little better, sacrificing a bit of quality.
Again, it would look miles better than FSR3.

So many engineers all wasted in this forum.

 

[winjer]

So many engineers all wasted in the forum.

Salty aren't you.

 

[Karim]

Salty aren't you.

Doesn't already some developer debunked it saying something like the FSR4 could take 14mss to the series X gpu or similar? That's the only concrete argument I found time ago about it to the base console.

 

[Fafalada]

It will only use a fraction of that per frame.

That's true for most compute workloads - the point isn't 'how many OPs/frame' - it's how fast the workload competes.
Eg. if a workload is 100% int-compute bound, and it completes in 1ms with 300TOP/s, a 50TOP/s will take 6ms for the same operation.
Both 'only' run for a fraction of the frame - but one is still 6 times longer.

Now how TOP compute bound these workloads are is another story. Case in point:

DP4A support on the PS5, we could have something like XeSS running on it

A hacked Linux PS5 with drivers from god knows what runs XeSS (and the Int8 path of all things) and apparently gets playable results.
Unfortunately DF decided not to actually measure its performance - tech channels eh

 

[winjer]

Doesn't already some developer debunked it saying something like the FSR4 could take 14mss to the series X gpu or similar?

No. That is just you making stuff up.
But several people on the Optiscaler Discord have showed FSR4 running on RDNA2 GPUs of similar performance.
Go there and read up, I can assure you you will learn a lot.

 

[Karim]

No. That is just you making studd up.
But several people on the Optiscaler Discord have showed FSR4 running on RDNA2 GPUs of similar performance.
Go there and read up, I can assure you you will learn a lot.

Doesn't RDNA2 on pc has also the famous infinite cache not available on consoles?

 

[winjer]

A hacked Linux PS5 with drivers from god knows what runs XeSS (and the Int8 path of all things) and apparently gets playable results.
Unfortunately DF decided not to actually measure its performance - tech channels eh

That was a terrible test by DF. The PS5 does not have support for DP4A, so it was probably using Linux emulation FP16 to Int8, which means a big loss in performance.

 

[Three]

Not this FSR debate again. Nobody asking why MS didn't do it already? Would have been a pretty good advantage to push, no?Especially as they have the Series S.

Every "it can" is not backed up by actual frametime costs and other drawbacks on fixed hardware with given framerate targets. It may be possible but it might not make sense for a lot of games. Didn't capcom even write their own upscaling algorithm for xbox? Why didn't those results blow us away?

 

[winjer]

Doesn't RDNA2 on pc has also the famous infinite cache not available on consoles?

ML uses mostly L2 and registers.
L3 helps, but it's not a deal breaker.
Even a 6600 with only 32MB and 128 bit bus can run XeSS DP4A very well. And even FSR4 Int8

 

[Karim]

ML uses mostly L2 and registers.
L3 helps, but it's not a deal breaker.
Even a 6600 with only 32MB and 128 bit bus can run XeSS DP4A very well. And even FSR4 Int8

And still we are in the hypothetical field. Too much convintion too many variables. Still I ask to you why higher AF are still so expensive on console but you are so sure AI upscaler will have like per like costs to the pc gpu.

 

[winjer]

And still we are in the hypothetical field. Too much convintion too many variables.

You can read up on Chips and Cheese, for the details on these GPUs. He does microbenching on all of these things.
Read up on RDNA2 and RDNA3 to learn how the caches work on these GPUs.

Chips and Cheese | Substack

The Devil is in the Details! Deep dives into computer hardware and software and the wider industry... Click to read Chips and Cheese, a Substack publication with thousands of subscribers.

chipsandcheese.com

 

[Karim]

You can read up on Chips and Cheese, for the details on these GPUs. He does microbenching on all of these things.
Read up on RDNA2 and RDNA3 to learn how the caches work on these GPUs.

Chips and Cheese | Substack

The Devil is in the Details! Deep dives into computer hardware and software and the wider industry... Click to read Chips and Cheese, a Substack publication with thousands of subscribers.

chipsandcheese.com

This doesn't change my doubts.

 

[winjer]

This doesn't change my doubts.

But it could change your ignorance. A lot.

 

[Karim]

But it could change your ignorance. A lot.

Listen take a bunch of lecture on the net not make you an expert of such stuff. So doesn't act like you are. I'm not exactly the only one who doubts about FSR4 feasible on console. Just sayin.

 

[Dorfdad]

XeSS is better than FSR3 and runs decently on RDNA2 and 3. So something like this could always be done, shame on MS for dropping their ML upscaler mid development.



Price also went up.

Not for those who already own one! Early adopters finally win one!!

 

[winjer]

Listen take a bunch of lecture on the net not make you an expert of such stuff. So doesn't act like you are.

I'm quoting experts. And pointing you to the people who know a lot more than both of us.
You can try to learn, or stay very ignorant.

 

[Karim]

I'm quoting experts. And pointing you to the people who know a lot more than both of us.
You can try to learn, or stay very ignorant.

The experts talk about the pc benchmarks. The only opinions I read concretely about the argument on the net not seem converge universally to the conclusion that FSR4 is easily "affordable" for something like the Series X console. PC equivalent asides.

 

[Bojji]

Not this FSR debate again. Nobody asking why MS didn't do it already? Would have been a pretty good advantage to push, no?Especially as they have the Series S.

Every "it can" is not backed up by actual frametime costs and other drawbacks on fixed hardware with given framerate targets. It may be possible but it might not make sense for a lot of games. Didn't capcom even write their own upscaling algorithm for xbox? Why didn't those results blow us away?

MS fucked up and neglected a lot of things this gen, them not creating ML upscaler for Xbox doesn't mean it couldn't be done.

Not for those who already own one! Early adopters finally win one!!

True. Today I saw Pro in store for like ~3900 PLN, I got mine for ~2600PLN over one year ago lol.

 

[Three]

MS fucked up and neglected a lot of things this gen, them not creating ML upscaler for Xbox doesn't mean it couldn't be done.

They were sending out teams to optimise other people's games/demos. You would think if they could get universal performance/IQ boosts across games they would take the opportunity. No?

Truth is nobody has really sat down and tried in practical terms to see the pros and cons of doing somthing like this and it's even likely that MS has. Just because you 'could' doesn't mean you 'should'. I've not seen a good case for 'should' yet, just that they could. Same with DF tests which is probably why they don't actually do any practical tests. Just mostly "look what is possible to run".

 

[Danjin44]

So did PS5 Pro finally became a Pro?

 

[welshrat]

Play cyberpunk on both and the difference is astounding, people pay 700 for that difference in GPU. Not coming from a console only perspective either as play a lot on good pc.

 

[Feel Like I'm On 42]

Play cyberpunk on both and the difference is astounding, people pay 700 for that difference in GPU. Not coming from a console only perspective either as play a lot on good pc.

Difference between what??

 

[Markio128]

So did PS5 Pro finally became a Pro?

It did. And Cerny is the pimp daddy.

 

[Clear]

The Pro's GPU is considerably bigger than that the Series X and benefits from a true bespoke ai upscaling implementation and people still downplay its advantages... So how's bodging FSR onto the Xbox going to be transformative ?

 

[Fafalada]

That was a terrible test by DF. The PS5 does not have support for DP4A, so it was probably using Linux emulation FP16 to Int8, which means a big loss in performance.

Not a Linux thing.
XeSS is an HLSL shader that will compile down to whatever the driver supports. Most likely individual int operations in this case (converting to FP would be incredibly cumbersome way to do this without a proper rewrite), so yes - it would not be using the hw effectively at all.
But the main point is we have no metric to argue how much it costs to run - just that it's 'useable' at 30fps and 1080p upscale from Switch resolutions in Cyberpunk.

 

[playbignbox]

No, and I think this kind of mid-gen update only makes it harder for Sony to encourage users to migrate to a new generation. There are many comments saying "we don't need a new generation yet" since many people find this hardware underutilized.

 

[winjer]

Not a Linux thing.
XeSS is an HLSL shader that will compile down to whatever the driver supports. Most likely individual int operations in this case (converting to FP would be incredibly cumbersome way to do this without a proper rewrite), so yes - it would not be using the hw effectively at all.
But the main point is we have no metric to argue how much it costs to run - just that it's 'useable' at 30fps and 1080p upscale from Switch resolutions in Cyberpunk.

I know that in Linux, it's possible to have Optiscaler use FP16 emulation to run Int8 XeSS.
As far as I know that is not possible in Windows. Though I could be wrong.

 

[Fafalada]

As far as I know that is not possible in Windows. Though I could be wrong.

It's an HLSL shader. You can compile it down to whatever you want, it's not an issue of emulation, or the OS.
Presumably just lack of windows driver support if it's not available.

I know that in Linux, it's possible to have Optiscaler use FP16 emulation to run Int8 XeSS.

Did anyone test perf on that?

 

[Brucey]

No, and I think this kind of mid-gen update only makes it harder for Sony to encourage users to migrate to a new generation. There are many comments saying "we don't need a new generation yet" since many people find this hardware underutilized.

Thats why ps6 isn't coming out anytime soon. PS5 Pro leading for years to come.

 

[dgrdsv]

This thread aged like milk

Not really. Still a completely useless upgrade for its price.

 

[Karim]

Thats why ps6 isn't coming out anytime soon. PS5 Pro leading for years to come.

2028 seems the only worst delay possible for the ps6 and still seems hazardous for sony

 

[winjer]

It's an HLSL shader. You can compile it down to whatever you want, it's not an issue of emulation, or the OS.
Presumably just lack of windows driver support if it's not available.

I'm not talking about compilation. I'm talking about running an FP8 FSR4 emulated into FP16, during runtime, using Optiscaler.

Did anyone test perf on that?

On Guru3d, there is a user named "TheCreator". He uses Linux, and sometimes runs tests with FSR4 FP8 emulated in FP16.
I remember him showing his test with FSR4 FP8 emulated on a 7900XTX, compared to FSR4 Int8 leaked dll. The FP16 was several frames slower.

 

[Karim]

Not really. Still a completely useless upgrade for its price.

Because...? The difference in the IQ is enormous. Also raytracing is a way easier to push on ps5 pro. Useless is surely an overdramatic expression to use.

 

[Bojji]

Because...? The difference in the IQ is enormous. Also raytracing is a way easier to push on ps5 pro. Useless is surely an overdramatic expression to use.

Difference is big only in games that actually use PSSR2 and RT. There are many games with barely any differences.

When you upgrade your GPU you get better results across the board. Being on the mercy of developers is the worst thing about consoles, thousands of games will be 30fps/1080p locked even on PS8.

 

[dgrdsv]

Because...? The difference in the IQ is enormous.

In a couple of titles. In most - you'd be hard pressed to see it from a typical viewing distance.

Also raytracing is a way easier to push on ps5 pro.

And this is resulting in what exactly? How many games "push RT on Pro" in a way which is visually transformative in comparison to the base unit?

 

[Skifi28]

Not really. Still a completely useless upgrade for its price.

Its high price scales with the price of the base console which also got higher and I imagine most people upgrading are trading their old one so they're not paying in full. Calling it "useless" when the difference in many games is night and day is pure ragebait.

In a couple of titles.

Right....

 

[dgrdsv]

Calling it "useless" when the difference in many games is night and day is pure ragebait.

How many? Are there even ten such games?

 

[Forsete]

In a couple of titles. In most - you'd be hard pressed to see it from a typical viewing distance.


And this is resulting in what exactly? How many games "push RT on Pro" in a way which is visually transformative in comparison to the base unit?

2 titles, really?