Yeah, more details are going to be needed on that front.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
[Insider Gaming] PS5 Pro Enhanced Requirements Detailed (mode that combines PSSR 4K upscaling, 60FPS, and RT, not a requirement for 'Enhanced' label)
- Thread starter Red Kong XIX
- Start date
We dont know whether the ML block is built into the GPU Compute Units or not.
Thing is, that the ML block's 300TOPs metric does not actually line up correctly with the rest of the PS5 Pro's GPU info (assuming the figure isnt rounded up from 268 but now it seems unlikely to me), so it woulf make sense if the ML block is not actually built into the CUs (but that would mean the APU is quite large).
The statement "You don't need the latest SDK for updating your games but you still get PSSR" sounds like a post-processing approach and nothing that is deeply embedded into the rendering pipeline. So yeah, it could be a hardware accelerator AFTER the classical pipeline.
I highly suspect the machine learning unit unit is going to be essentially the tempest engine 2.0.
Sony already have an area of custom CUs in PS5 for audio processing and the verge have also noted that the audio processing in PS5 Pro is better.
So they probably expanded the tempest engine from the 2CUs in PS5 and have modified them further to be better machine learning units.
Total speculation of course but it would line up with everything we know. Better audio, really really strong machine learning performance that does not make sense with our understanding of GPU CUs. PSSR can be added to games without even updating the SDK (render pipeline seemingly unaffected).
EDIT: didn't cerny insinuate when announcing the PS5 hardware that the tempest engine CUs could be used for more than just audio or am I misremembering? Maybe hinting at PSSR?
Last edited:
Is that how it works with DLSS as well? I remember at some point NVidia added a similar parallel computatiom approach.
He mentioned, "other algorithms that are either compuationally expensive, or need high bandwidth."
So, interestingly, the concept of doing DLSS upscaling of one frame concurrently with the rendering of the next frame was confirmed as part of the Ampere featureset by NVIDIA back in 2020, but it's never been used on PC, probably because the cost is frequently too minor for the devs to bother with it, and weaker cards that take longer to upscale will often target lower resolutions anyway. In a console, though, especially one always targeting 4K, it makes way more sense. I'm hoping we see it in Switch 2 as well. We don't know if this concurrency is also a PSSR feature, but I would like to hope that it's possible. An extra 2ms of input lag isn't the worst thing in the world - especially if you get a better framerate that lowers the lag in turn.
2ms, even for people obsessed with lowering input lag is nothing for the benefit of much better fps.
Bingo - I remember reading that God of War 2018 running at 60fps on PS5 has something like 120ms. This whole subject came up on some discussions about the Switch 2 where the cost of DLSS 2160p could be like 10-15ms - awful as part of the frame render but much more viable as input lag. The PS5 Pro wouldn't need it as much, but it would absolutely still benefit, especially when trying to hit 120fps.
Yeah sorry, I mixed those upscalers with the frame generation stuff. For the (spatial) upscalers they use a trained autoencoder (16K/8K/4K -> Native resolution -> Target resolution), put the kernel it the graphics drivers and just apply the multilayer filtering (CNN) to reconstruct the image.
For the frame-generation stuff they need motion vectors, etc. to interpolate the inter-frames correctly, therefore you need information extracted within the rendering pipeline.
PlayStation 5 uncovered: the Mark Cerny tech deep dive
On March 18th, Sony finally broke cover with in-depth information on the technical make-up of PlayStation 5. Expanding …
www.eurogamer.net
"The Tempest engine itself is, as Cerny explained in his presentation, a revamped AMD compute unit, which runs at the GPU's frequency and delivers 64 flops per cycle. Peak performance from the engine is therefore in the region of 100 gigaflops, in the ballpark of the entire eight-core Jaguar CPU cluster used in PlayStation 4. While based on GPU architecture, utilisation is very, very different.
"GPUs process hundreds or even thousands of wavefronts; the Tempest engine supports two," explains Mark Cerny. "One wavefront is for the 3D audio and other system functionality, and one is for the game. Bandwidth-wise, the Tempest engine can use over 20GB/s, but we have to be a little careful because we don't want the audio to take a notch out of the graphics processing. If the audio processing uses too much bandwidth, that can have a deleterious effect if the graphics processing happens to want to saturate the system bandwidth at the same time."
Essentially, the GPU is based on the principle of parallelism - the idea of running many tasks (or waves) simultaneously. The Tempest engine is much more serial-like in nature, meaning that there's no need for attached memory caches. "When using the Tempest engine, we DMA in the data, we process it, and we DMA it back out again; this is exactly what happens on the SPUs on PlayStation 3," Cerny adds. "It's a very different model from what the GPU does; the GPU has caches, which are wonderful in some ways but also can result in stalling when it is waiting for the cache line to get filled. GPUs also have stalls for other reasons, there are many stages in a GPU pipeline and each stage needs to supply the next. As a result, with the GPU if you're getting 40 per cent VALU utilisation, you're doing pretty damn well. By contrast, with the Tempest engine and its asynchronous DMA model, the target is to achieve 100 percent VALU utilisation in key pieces of code." "
Sorry for long quote, I thought it all might be relevant .
PlayStation 5 uncovered: the Mark Cerny tech deep dive
On March 18th, Sony finally broke cover with in-depth information on the technical make-up of PlayStation 5. Expanding …
"The Tempest engine itself is, as Cerny explained in his presentation, a revamped AMD compute unit, which runs at the GPU's frequency and delivers 64 flops per cycle. Peak performance from the engine is therefore in the region of 100 gigaflops, in the ballpark of the entire eight-core Jaguar CPU cluster used in PlayStation 4. While based on GPU architecture, utilisation is very, very different.
"GPUs process hundreds or even thousands of wavefronts; the Tempest engine supports two," explains Mark Cerny. "One wavefront is for the 3D audio and other system functionality, and one is for the game. Bandwidth-wise, the Tempest engine can use over 20GB/s, but we have to be a little careful because we don't want the audio to take a notch out of the graphics processing. If the audio processing uses too much bandwidth, that can have a deleterious effect if the graphics processing happens to want to saturate the system bandwidth at the same time."
Essentially, the GPU is based on the principle of parallelism - the idea of running many tasks (or waves) simultaneously. The Tempest engine is much more serial-like in nature, meaning that there's no need for attached memory caches. "When using the Tempest engine, we DMA in the data, we process it, and we DMA it back out again; this is exactly what happens on the SPUs on PlayStation 3," Cerny adds. "It's a very different model from what the GPU does; the GPU has caches, which are wonderful in some ways but also can result in stalling when it is waiting for the cache line to get filled. GPUs also have stalls for other reasons, there are many stages in a GPU pipeline and each stage needs to supply the next. As a result, with the GPU if you're getting 40 per cent VALU utilisation, you're doing pretty damn well. By contrast, with the Tempest engine and its asynchronous DMA model, the target is to achieve 100 percent VALU utilisation in key pieces of code." "
Sorry for long quote, I thought it all might be relevant .
I also thought it a bit strange that Sony would spend so much engineering time on audio. Developing their own bespoke variant of CUs seems like a lot of work.
Now if this was the initial work for PSSR that makes a whole lot more sense. You shed your audio from the CPU (which did it in PS4) and you shed your reconstruction from the GPU (which does it in PS4 and PS5).
I might be well off base but I think it would be most logical for PSSR to be handled by an expanded tempest engine. Maybe Sony started developing it because they saw on the AMD roadmap they wouldn't have AI reconstruction so worked on their own solution.
not sure why people only talk about 30 or 60 fps, when there is a perfectly viable happy middleground in 40fps
if base PS5 only delivers 30fps, there is a good chance PS5 pro can do 40fps
if base PS5 only delivers 30fps, there is a good chance PS5 pro can do 40fps
Basically we just need to get to a point where TVs that support 40FPS have been widely adopted. Honestly not sure where we are at with that, but I personally still have my old Plasma TV, so no shiny new refresh rates/40FPS for me!
It would be great for games to focus on 40FPS as an additional target though. Obviously I dont wanna see 60FPS dropped, but its a great middle ground between graphics and smoother performance.
NVidia uses motion vectors for DLSS (2+), its a temporal upscaler, PSSR will be too.
If its made to be serializable then it wouldnt work well as a ML hardware. Deep Learning models do Matrix Multiplications and those run well with high parallelization.PlayStation 5 uncovered: the Mark Cerny tech deep dive
On March 18th, Sony finally broke cover with in-depth information on the technical make-up of PlayStation 5. Expanding …
"The Tempest engine itself is, as Cerny explained in his presentation, a revamped AMD compute unit, which runs at the GPU's frequency and delivers 64 flops per cycle. Peak performance from the engine is therefore in the region of 100 gigaflops, in the ballpark of the entire eight-core Jaguar CPU cluster used in PlayStation 4. While based on GPU architecture, utilisation is very, very different.
"GPUs process hundreds or even thousands of wavefronts; the Tempest engine supports two," explains Mark Cerny. "One wavefront is for the 3D audio and other system functionality, and one is for the game. Bandwidth-wise, the Tempest engine can use over 20GB/s, but we have to be a little careful because we don't want the audio to take a notch out of the graphics processing. If the audio processing uses too much bandwidth, that can have a deleterious effect if the graphics processing happens to want to saturate the system bandwidth at the same time."
Essentially, the GPU is based on the principle of parallelism - the idea of running many tasks (or waves) simultaneously. The Tempest engine is much more serial-like in nature, meaning that there's no need for attached memory caches. "When using the Tempest engine, we DMA in the data, we process it, and we DMA it back out again; this is exactly what happens on the SPUs on PlayStation 3," Cerny adds. "It's a very different model from what the GPU does; the GPU has caches, which are wonderful in some ways but also can result in stalling when it is waiting for the cache line to get filled. GPUs also have stalls for other reasons, there are many stages in a GPU pipeline and each stage needs to supply the next. As a result, with the GPU if you're getting 40 per cent VALU utilisation, you're doing pretty damn well. By contrast, with the Tempest engine and its asynchronous DMA model, the target is to achieve 100 percent VALU utilisation in key pieces of code." "
Sorry for long quote, I thought it all might be relevant .
Because 40fps 120hz modes have been around for years now and the vast majority of developers still seemingly refuse to use them, even when they would be tremendously valuable (see various recent Final Fantasy and UE5 titles). 30 and 60fps modes remain the common standard, including in every game with a 40fps mode, so obviously we're going to focus on them - even if I wish 40fps was used more. One reason why I wanted a bigger CPU upgrade was so the Pro could guarantee a 40fps mode in almost any game would at least be possible, but oh well.
because nobody has a TV capable of displaying 40fps, everyone has a TV that can do 30 or 60
A 40fps mode in the vast majority of games is possible now without a cpu upgrade…Nothing would ever be guaranteed with devs, even with the cpu upgrade. But Sony devs have done a great job this gen doing it so far.
Which is sad, 40hz modes are the absolute best. Props to Sony for supporting them in most of their games.
This is a massive exaggeration in the other direction. 120hz TVs have been around for years, millions of people own then. Even just LG OLED TVs like the CX I got 3 years ago have sold in the multiple millions, and they target exactly the sorts of enthusiasts who buy consoles early.
In the vast majority of games, yes, but I would've liked to see a CPU boost of about 33%, so that even in the worst-case scenario, a hypothetical PS5 game that was CPU-bound at 30fps could do 40 on the Pro. I guess we'll just have to hope that devs try to do more 40fps modes on the base consoles.
Julien Chièze said he talked to many devs about the PS5 Pro who confirmed thinsg to him and eh also saw things about it:
View: https://www.youtube.com/watch?v=_NabkE8vdj0
View: https://www.youtube.com/watch?v=_NabkE8vdj0
Well the weird thing about RDNA 3 is that it features tensor-like cores, I think the full 7900 XTX has 192 of them. But they don't do anything currently. Maybe a future version of FSR could make use of them, but I think they're just sitting there inactive for now.
If by "poorly optimized" games you mean games which are aiming at 30 fps by design then yeah, these will "struggle" with going above that, and their number will be rising.
There is no downplaying of anything.
This generation has gotten a lot of 60 fps games so far because it got a) a very long cross-gen period with games being designed for PS4/XBO (or even Switch) baselines and b) better resolution reconstruction approaches which allow somewhat okayish reconstruction from down to 1/4th of display resolution.
Both of these are out of the picture once we're past cross-gen as both depend on the baseline h/w target. Expecting everyone to just continue targeting PS4/XBO graphics for some reason is what's "downplaying reality" here.
We are seeing it because of h/w targets and when we get a game which didn't target XBO we're not getting anything close to a stable usable 60 either.
From my casual viewing of DF analysis over the last year or so it seem like the vast majority of current gen releases aren't hitting anything close to a locked stable 60 anymore.
Yep, same reason: these games are/were made for previous gen h/w.
Can you summarize, not many people on this forum speak French.
Having PSSSR being better than FSR 2 is not a fetat at all to brag about. Everythign is better than FSR 2 so if they claim victory for being better than the worst thing ever created they really should push the things much much more before release and not rely on future iterations of PSSSR to solve things out.
Last edited:
I hope PSSR can be around what the 1.3 version of XeSS on the XMX path in terms of quality. Big reduction of ghosting, nice temporal stability, and significantly fewer artifacts on water and still with a nice, clean overall IQ
It hits all of the things that have been standing out to me in many recent games
It hits all of the things that have been standing out to me in many recent games
Not much more than what i said. He is just discussing the the verge article talking about the new bosot mode but he said he spoke to many devs from many studios who showed him things (mostly games) and he insinuated that you will be surprised. I am sure he can't say more cus he may fear the MLID case which he discussed too about nuking his video by copyright so I understand why he can't say more.
Edit: my personal speculation is that he at least saw stuff from Ubisoft cus he is buddy with some of them and you know Ubi is mostly a french studio eventhough it has many branches allover the wolrd but it is easier to get stuff from a French studio while you are in France and have connections. But he said many companies so he saw more.
Reason that they compare to FSR2 is because thats the main 3rd party alternative on PS. Its pointless to compare to DLSS because PS5 does not support it.
It is basically telling developers it is a better option than what they previously had on the console.
Last edited:
Yes I can understand that but how much it is better it remains to be seen.
Last edited:
The level PSSR needs to reach is "good enough". The main issue are games like Rebirth where you immediately can tell the big hit to the IQ without pixel counting. If you can reach the level of, say, GoW Ragnarok perf mode where you most of the time cannot even tell the difference between perf and quality mode IQ, you got yourself a win.
I don't know if this is directly related to the PS5 Pro but it started that way anyway:
View: https://twitter.com/dark1x/status/1780341879919288359
Kepler faslely assumed that coding to the metal doesn't happen anymore and that optimisations are things of the past but he got halted by our John who witnessed insane stuff. I wonder if coding to the metal he saw was related to the PS5 Pro or in a certain game in general.
View: https://twitter.com/dark1x/status/1780341879919288359
Kepler faslely assumed that coding to the metal doesn't happen anymore and that optimisations are things of the past but he got halted by our John who witnessed insane stuff. I wonder if coding to the metal he saw was related to the PS5 Pro or in a certain game in general.
Also, I think an argument can be made that an existing PS5 owner that doesn't have a 4K HDR 120Hz set would get far more benefit out of upgrading their TV rather than upgrading to a PS5 Pro.
Sounds like a game, and does seem like console related at least.
View: https://twitter.com/dark1x/status/1780356333142786266
Indeed - the stills look good but then stills are most flattering when the main problem with AI reconstruction is fast movement.
Good thing is that they can update it over time, as we have seen with DLSS and XeSS. How it is on launch will be the floor for quality.
There are plenty of games that are current gen only, that are a generally solid 60fps. Frame drops here and there still happened on cross gen games, not major issues like say Immortals.
What worries me is the launch iteration even the document mentions that lot of improvements are coming in later versions like support for upsclaed 8K and more optimisations. If launch version is just slightly improved over crappy FSR 2 and much behind let's say DLSS 2.0, we are scerwed, cus even FSR 3.1 is supposed to improve things aand is aroudn the corner and FSR 4.0 will happen mostly before the PSSSR version getting updated and you know how many games are hungry for such upsacling tech to off load the burdens of Rt and performance to it. So if the launch version is meh, we may not expect the games who will support to get updated each time with another version each time so we will be stuck with a FSR 2.3 like(i Know I am exaggerating but still, if this happens we won't be that far away).
Surprised meaning that the improvements are better than specs indicate?
PSSR is machine learning which should put it above fsr 2, which isn't.
I think that's AMD's biggest issue. Though, I'm no expert.
We actually do not have the whole picture of the GPU and features even with the spec leaks.
The only reason this much effort would go into PSSR, would be if they had a solution that can perform and look better than FSR by a good margin. This is costing them millions in hardware and software development so it better be.
Yeah, if FSR was good enough they wouldn't have gone out of there way to do something unique.
I am pretty sure FSR 4.0 will be a decent upgrade. AMD will surely not stay there being frozen unless they enjoy the backlash.
That would be funny if devs have to choose between FS 4 and PSSSSR in the future. XD
And as a result of how closely the work with AMD on the APU, all of this work on PSSR and RT hardware will feed back to AMD's own hardware.
FSR 4 might include similar dedicated hardware support as PSSR.
Well we know for a fact that Naughty dog were doing insane stuff (literally coding to the metal) with PS3 hardware in TLOU.
Yes theoretically developers can already use one whole thread (half of the Tempest CU?) on any tasks, even graphical ones. Like once I thought it could be used to offload some RT related tasks. Ideally tasks that would work very well with the DMA model (no cache, direct access to main memory).PlayStation 5 uncovered: the Mark Cerny tech deep dive
On March 18th, Sony finally broke cover with in-depth information on the technical make-up of PlayStation 5. Expanding …
"The Tempest engine itself is, as Cerny explained in his presentation, a revamped AMD compute unit, which runs at the GPU's frequency and delivers 64 flops per cycle. Peak performance from the engine is therefore in the region of 100 gigaflops, in the ballpark of the entire eight-core Jaguar CPU cluster used in PlayStation 4. While based on GPU architecture, utilisation is very, very different.
"GPUs process hundreds or even thousands of wavefronts; the Tempest engine supports two," explains Mark Cerny. "One wavefront is for the 3D audio and other system functionality, and one is for the game. Bandwidth-wise, the Tempest engine can use over 20GB/s, but we have to be a little careful because we don't want the audio to take a notch out of the graphics processing. If the audio processing uses too much bandwidth, that can have a deleterious effect if the graphics processing happens to want to saturate the system bandwidth at the same time."
Essentially, the GPU is based on the principle of parallelism - the idea of running many tasks (or waves) simultaneously. The Tempest engine is much more serial-like in nature, meaning that there's no need for attached memory caches. "When using the Tempest engine, we DMA in the data, we process it, and we DMA it back out again; this is exactly what happens on the SPUs on PlayStation 3," Cerny adds. "It's a very different model from what the GPU does; the GPU has caches, which are wonderful in some ways but also can result in stalling when it is waiting for the cache line to get filled. GPUs also have stalls for other reasons, there are many stages in a GPU pipeline and each stage needs to supply the next. As a result, with the GPU if you're getting 40 per cent VALU utilisation, you're doing pretty damn well. By contrast, with the Tempest engine and its asynchronous DMA model, the target is to achieve 100 percent VALU utilisation in key pieces of code." "
Sorry for long quote, I thought it all might be relevant .
But I don't believe in them using tempest engine 2 for int-8 processing (and the CUs would need too many modifications, essentially becoming tensor cores). I think it would be simpler for them to use the new RDNA4 WMMA units. This is the kind of thing Sony would absolutely do: use features not yet released on PC ecosystem. They did it with PS4 Pro. And they should also do it, allegedly, with new RDNA4 RT hardware.
View: https://twitter.com/Kepler_L2/status/1768590437206339657
When is FSR 4 coming? Two years? We haven't even seen the new FSR 3 released that they just previewed with those ratchet screens.
Its also work that will probably be used in the inevitable PS6.
Even more likely if the same or similar solution is being used by AMD, they will have a clear development path for the tech.
And the compromises they keep making to hit 60 FPS are getting really rough. Like the 432p on aveum, or bizzarely using FSR1 or freaking nearest neighbor upscaling from 720p with FF
Of course that these kinds of solutions will get refined more as time goes on.
The secret with deep learning is no one actually has a clue why their solution "worked" (kinda), because those models are really complicated, improvements from where the faults are found would be a constant attempt at making up heuristics to the problem as well as trial and error until it works well. They won't release a solution that suddenly blows everything else out of the water unless if they have something really novel going on here (could be the spectral transformation I wondered about, but wouldn't bet on it alone being some silver bullet), as well as being very lucky.
Kepler is probably wrong about that, the documents say the hardware solution is fully custom, not semi custom (which would be the case if they just back ported an AMD feature). They have designed their own solution to ML acceleration.
The secret with deep learning is no one actually has a clue why their solution "worked" (kinda), because those models are really complicated, improvements from where the faults are found would be a constant attempt at making up heuristics to the problem as well as trial and error until it works well. They won't release a solution that suddenly blows everything else out of the water unless if they have something really novel going on here (could be the spectral transformation I wondered about, but wouldn't bet on it alone being some silver bullet), as well as being very lucky.
Kepler is probably wrong about that, the documents say the hardware solution is fully custom, not semi custom (which would be the case if they just back ported an AMD feature). They have designed their own solution to ML acceleration.
Although the people preferring performance mode will likely still choose it for solid 60fps even with reduced image quality, there are plenty of current gen only games that do not suffer from this. At least they are using upscaling that is better than the FSR 1 or whichever technique was in FF7R-2.