Sony PlayStation 5 Pro

[Remij]

Who cares about all that? Let's just compare results, shall we?

No.. we wont just compare results.. we will discuss, observe, and make guesses as to how it works on the hardware alongside that. If you can't handle people suggesting that PSSR is running on the shader cores and not dedicated cores, then you should take your own advice instead of arguing with people about it..

This is literally the forum to discuss and theorize about how PS5 Pro works on a technical level.

 

[Cynicalking]

I find it hilarious that in a discussion forum no discussion can be had without always the same people getting emotional over brand preference.

 

[techuse]

Some people want to understand the how and why behind these technologies. Us discussing it doesn’t affect your enjoyment of the end result.

 

[Charlietus]

Again, the mere *existence* of options creates FOMO. That's the whole psychological issue to begin with that makes the PS5 Pro potentially worthwhile. But if you're pushing options on PS5 Pro users as well, then you're not getting rid of the FOMO.

If a console was a milion teraflops supercomputer, developers would probably still offer options for the framerate

You can always do double at 33.3 ms.

 

[Shifty Geezer]

Who cares about all that?

Everyone trying to talk about it. If that's not a topic that interests you, just leave everyone else to it.

 

[Shifty Geezer]

Some people want to understand the how and why behind these technologies. Us discussing it doesn’t affect your enjoyment of the end result.

Pretty much everyone on this forum should be wanting to understand the how and why behind these technologies - it's B3D's reason d'etre!

 

[pjbliverpool]

Right now PSSR is been compared to be roughly at the same level as DLSS 3.7 and the console is not even released.

From this DF clip comparing PSSR against all others solution: FSR3.1, XeSS 1.3 and DLSS 3.7 using arguably the most difficult use case for reconstruction techniques: the dreaded grass!

DLSS = PSSR > XeSS >>> FSR3.1


In motion we can see why PSSR is superior to XeSS.

You keep making the claim that PSSR is equivalent to DLSS but that is not at all the conclusion of the comparison video's you're referencing. PSSR does improve upon DLSS's output in one or two areas in R&C (namely RT checkerboard reconstruction and edge aliasing on some curved surfaces), but is more generally less stable, softer, and suffers from other issues like delayed RT reflections. And key to the comparison as pointed out by Alex is that the PS5 Pro instance is actually implemented in a superior way to the PC instance because it uses a higher sample input on the CB RT reflections, (hardly surprising given it's intended use as a launch showpiece for the tech, and implemented by a Sony studio), hence why those CB reflections look better on the Pro. This doesn't have to be the case with DLSS.

Similarly, the XeSS comparison is from one single scene/use case and yes while PSSR wins out in that single comparison, that is far too little data to make sweeping conclusion's like "it's superior to XeSS".

With regards to the point around DLSS being at v3.7 and PSSR 'not even launched'. I don't really see why that is relevant. Just because it only just launched doesn't mean the model hasn't been being worked on for the last several years. I'd be very surprised if PSSR doesn't share a heritage with the upcoming FSR4 AI based upscaler from AMD and it seems likely that AMD would have started working on their own AI based upscaler not long after DLSS originally debuted 6 years ago, likely waiting to launch it until it was at a quality level sufficient to compete reasonably with DLSS and the hardware side was in place with RDNA4. XeSS was also far better than DLSS 1.0 (or 1.9) when it first released but no-one is claiming it's an inherently better model as a result.

 

[DavidGraham]

Yes. However, if they were only doing upscaling, the rest of the time they'd be dead.

I am genuinely curious to know why? as we all know GPUs work through pipelining, the ML units will just be a part of the pipeline and would be busy doing work all the time, just like Texture units, Raster units, RT units, FP units .. etc. In fact, this is the case for Tensor units in RTX hardware already. In fact, none of the various units in the GPU work for the whole 16ms of the frame, Texture units would also finish their work in ~2ms or 3ms and pass their work to some other unit and then receive more work from other units, same thing happens with FP units .. etc.

5Pro's description of its ML abilities is upscaling only

Well, If Sony is following in NVIDIA steps, then I believe frame generation and denoising are coming, but these things take time to train the models, as such they will be revealed when they are ready.

 

[Shifty Geezer]

I am genuinely curious to know why?

It's part of the wider conversation. Really the only bit that needs thinking about is the second part of that post, the three options.

Well, If Sony is following in NVIDIA steps, then I believe frame generation and denoising are coming, but these things take time to train the models, as such they will be revealed when they are ready.

But why not say that?! It's crazy to have that potential and a price tag everyone balked at and not market it as "advanced ML that'll bring great new features, starting with PSSR out of the gate." No?

Now if the ML hardware is part of the CUs and ties the CUs up when in operation, then every ms of ML activity mean one less ms of other CU activities, which means losing rendering time. So yes, as you say, ML hardware that's only used 2ms of the frametime makes sense, like texture units only used some of the time. That's the argument being made. The hardware is there as part of RDNA 4 (or whatever number) but it can't really be used for more than 2ms or else there's just not enough rendering power. It's not 'discrete' (independent, whatever, the word choice keeps getting in the way!) ML hardware that can operate tangentially to the CUs with the option for 16 ms of ML activity per frame alongside 16 ms of compute rendering work.

What's your take on that assessment?

 

[Charlietus]

It's part of the wider conversation. Really the only bit that needs thinking about is the second part of that post, the three options.

But why not say that?! It's crazy to have that potential and a price tag everyone balked at and not market it as "advanced ML that'll bring great new features, starting with PSSR out of the gate." No?

They aren't going to promise features that could come in a year or more (if ever, maybe they are keeping them for PS6...)

It's like if Nvidia said that ray reconstruction or AI HDR was in the works at the 3000 series launch.

Those companies try to study the most effective way and timing to announce things, and announcing them years before is not how you do it.

In PS5 Pro's case, maybe it could do all three features at the same time, but unless Sony announces anything, I wouldn't hope to get anything more than PSSR.

(Another aspect to consider is how much memory DLSS frame generation consumes, and it's a lot.)

 

[Remij]

They aren't going to promise features that could come in a year or more (if ever, maybe they are keeping them for PS6...)

It's like if Nvidia said that ray reconstruction or AI HDR was in the works at the 3000 series launch.

Those companies try to study the most effective way and timing to announce things, and announcing them years before is not how you do it.

In PS5 Pro's case, maybe it could do all three features at the same time, but unless Sony announces anything, I wouldn't hope to get anything more than PSSR.

(Another aspect to consider is how much memory DLSS frame generation consumes, and it's a lot.)

So let me ask you.... Why exactly does discussing and speculating on this question make you upset? The entire point is to question these things, and like Shifty said, if the hardware is capable of it they would make it known. And since they aren't saying anything, people are free to speculate all they want, and given some of the decisions being made by some games, it leads in a certain direction... until Sony, or some other 3rd party come and clear it up.

Like, I don't get it. People are suggesting that the hardware likely isn't capable of it, and you're getting annoyed by it.. while at the same time telling us "I wouldn't hope to get anything more than PSSR"... so are you agreeing with us or not?

 

[iroboto]

Well, If Sony is following in NVIDIA steps, then I believe frame generation and denoising are coming, but these things take time to train the models, as such they will be revealed when they are ready.

One of the nice things about a dedicated core like tensor cores, is that they can be used for frame generation and denoising for the in-between frame, while the SIMT units and the rest of the GPU are working on the current frame.

 

[Charlietus]

So let me ask you.... Why exactly does discussing and speculating on this question make you upset? The entire point is to question these things.. and like Shifty said... if the hardware is capable of it... they would make it known. And since they aren't saying anything.. people are free to speculate all they want, and given some of the decisions being made by some games, it leads in a certain direction... until Sony, or some other 3rd party come and clear it up.

Like.. I don't get it... people are suggesting that the hardware likely isn't capable of it, and you're getting annoyed by it.. while at the same time telling us "I wouldn't hope to get anything more than PSSR"... so are you agreeing with us or not?

Where is my post am I coming off as annoyed? I'm as cool as a cocomber, Just discussing why I think they wouldn't talk about some things now.

As far as if I'm agreeing or not, I'm a bit in the middle. I think the hardware is capable of doing more features, but at the same time, either:

•they are not ready

•memory isn't enough

•they are keeping them for PS6

•a bit of the three combined

 

[Remij]

Where is my post am I coming off as annoyed? I'm as cool as a cocomber, Just discussing why I think they wouldn't talk about some things now.

As far as if I'm agreeing or not, I'm a bit in the middle. I think the hardware is capable of doing more features, but at the same time, either:

•they are not ready

•memory isn't enough

•they are keeping them for PS6

•a bit of the three combined

I apologize, I mixed you and Globalisateur up!

 

[Shifty Geezer]

They aren't going to promise features that could come in a year or more (if ever, maybe they are keeping them for PS6...)

Um...that's exactly what console companies have done for generations! Often making promises that are never delivered.

It's like if Nvidia said that ray reconstruction or AI HDR was in the works at the 3000 series launch.

Somewhat different is back then, they didn't know what ML could do. Now we do.

 

[Shifty Geezer]

The entire point is to question these things, and like Shifty said, if the hardware is capable of it they would make it known.

To be clear, they might not in a confusing move that makes no sense (to me). I'm not trying to assert a position as the only truth. The speculation and seeing if you are right or not is the game! It's a logic puzzle, with missing pieces. Can we identify reality before its revealed?

For me, the logic is clear, and it'd require an illogical piece to be in play for there to be more to it than my theory. That doesn't make that impossible. My virtual five pound bet is on the ML being part of the CUs such that they are occupied when running ML workloads. Hopefully one day we'll learn exactly what's happening and we see whether I win that virtual bet or not.

Importantly, none of this matters! No-one should be getting upset!!! Emotions really shouldn't be coming into it except a mild joy from the logical challenge.

 

[DavidGraham]

Now if the ML hardware is part of the CUs and ties the CUs up when in operation, then every ms of ML activity mean one less ms of other CU activities, which means losing rendering time

Yeah indeed, if the ML units are integtated within the CUs (as in just accelerated instructions similar to the rapid packed math FP16x2), then yes, any ML workload is going to encroach on FP32 throughput, reducing it substantially, especially if we account for things like register pressure and memory bandwidth.

This would be the same situation as RT units in current consoles and current AMD GPUs, they are integtated within the Texture units, which is why RT is weak on these GPUs, texturing reduce RT throughput and vice versa, ML units being structured in similar way would lead to the same result.

If the PS5 Pro is working like this, then they won't afford to do anything ML other than upscaling, there is simply no hardware throughput to do anything else. The PS5 Pro would also be limited in the amount of upscaling it can do. DLSS for example can afford to do competent 8X upscaling (from 1080p to 8K), while other upscalers can't. DLSS can also handle competent ML Anti Aliasing at native resolution (DLAA) without too much of a performance impact, and it can do ML downsampling (DLDSR) from any resolution. These things would be outside the scope of PS5 Pro ML capabilities.

It would also mean Sony calculated how much ML it exactly needed for upscaling, without an ounce more.

I guess we should be on the lookout for upcoming implementations, if we see PS5 Pro games with nothing more than typical 4x PSSR upscaling, if we see some heavy games forgo PSSR completely in favor of lighter upscaling algorithms, then that would be a very powerful indication of the limited integrated ML capabilities of the PS5 Pro.

However, if we see games with with further ML capabilities like PSSR 8x upscaling to 8K or games doing multiple ML algorithms at the same time like, PSSR frame generation or PSSR AA, or HDR conversion, then that would be evidence that PS5 Pro possess more advanced discrete ML capabilities.

 

[snc]

Yeah indeed, if the ML units are integtated within the CUs (as in just accelerated instructions similar to the rapid packed math FP16x2), then yes, any ML workload is going to encroach on FP32 throughput, reducing it substantially, especially if we account for things like register pressure and memory bandwidth.

This would be the same situation as RT units in current consoles and current AMD GPUs, they are integtated within the Texture units, which is why RT is weak on these GPUs, texturing reduce RT throughput and vice versa, ML units being structured in similar way would lead to the same result.

If the PS5 Pro is working like this, then they won't afford to do anything ML other than upscaling, there is simply no hardware throughput to do anything else. The PS5 Pro would also be limited in the amount of upscaling it can do. DLSS for example can afford to do competent 8X upscaling (from 1080p to 8K), while other upscalers can't. DLSS can also handle competent ML antialiasing at native resolution (DLAA) without too much of a performance impact, and it can do ML downsampling from any resolution. These things would be outside the scope of PS5 Pro ML capabilities.

It would also mean Sony calculated how much ML it exactly needed for upscaling, without an ounce more.

I guess we should be on the lookout for upcoming implementations, if we see PS5 Pro games with nothing more than typical 4x PSSR upscaling, then that would be a very powerful indication of the limited integrated ML capabilities of the PS5 Pro.

However, if we see games with with further ML capabilities like PSSR 8x upscaling to 8K or games doing multiple ML algorithms at the same time like, PSSR frame generation or PSSR AA, or HDR conversion, then that would be evidence that PS5 Pro possess more advanced discrete ML capabilities.

Alan wake 2 is upscaling to 4k from 864p

 

[RobertR1]

Who cares about all that?

People participating in the technology subsection of the console forum? Why not just make a general thread about pssr if you’re getting worked up about people wanting to discuss the underlying technology and not just output comparisons?

 

[Karamazov]

Do DF already have a unit ?