Support NeoGAF

[SlimySnake]

Just saw this gaming tech video. Pros performance mode runs at 1440p internal resolution just like base PS5s performance mode. PSSR then reconstructs the image to 4k. Looks way better than 1440p.

However, the framerate comparison could help us determine just how much the PSSR reconstruction could cost on the GPU. The framerates are roughly in line on both the Pro and PS5 suggesting most of the circa 45% extra horsepower is going into PSSR reconstruction. In other Pro games, we have seen an extra 30% raw gpu performance so it could also be just 30%. In some cases, it dips around 10 fps below the ps5 performance mode like screenshotted below but that could also be a vram bottleneck since the textures and other effects are still rendered at 4k.

Timestamped:

Regardless, the reconstruction costs are in line with what we've seen from DLSS and FSR.

 

[Calverz]

Looking forward to this on pc. Il play it ultrawide 100+ fps and 4k like I did the first game.

 

[Gaiff]

He also compares TAA to PSSR. However, I'm not sure if TAA and PSSR are running the same input and output resolutions.

 

[rofif]

We don't really know what's going on under the hood

 

[Bojji]

We don't really know what's going on under the hood

We know last of us res is 1440p, just like PS5 version.

 

[OverHeat]

For Sony games on PC PSSR will come too I’m sure.

 

[Bojji]

For Sony games on PC PSSR will come too I’m sure.

On AMD GPUs?

I doubt pssr tech will leave PS5 pro.

 

[Gaiff]

For Sony games on PC PSSR will come too I’m sure.

Maybe...but why waste their time when there's already FSR, DLSS, XeSS, and TAA? Adding a 5th solution won't be providing much.

 

[zeroluck]

Assuming 35% faster gpu and no cpu bottleneck, ps5 pro without pssr should be getting 111 * 1.35 = 150 or 6.66ms frametime, with pssr its getting 102 or 9.8ms, so pssr takes 3.14ms, that is actually way slower than what I expected, a 3080 takes ~1ms to upscale from 1080p to 4k.

 

[Unknown?]

Looking forward to this on pc. Il play it ultrawide 100+ fps and 4k like I did the first game.

Weak sauce. You need 1000fps to make the motion clarity as good as a CRT and 4k is the 720p of 2025. 8k is the future and soon we will be using 16k.

 

[DeathGuise]

Weak sauce. You need 1000fps to make the motion clarity as good as a CRT and 4k is the 720p of 2025. 8k is the future and soon we will be using 16k.

You just killed my brain with the continuity of this. Well done

 

[SlimySnake]

Assuming 35% faster gpu and no cpu bottleneck, ps5 pro without pssr should be getting 111 * 1.35 = 150 or 6.66ms frametime, with pssr its getting 102 or 9.8ms, so pssr takes 3.14ms, that is actually way slower than what I expected, a 3080 takes ~1ms to upscale from 1080p to 4k.

Nah. ive done several tests on this in several different games and it varies from game to game. some games cost 20% while others can go up as high as 35%.

 

[winjer]

I bet the performance cost of PSSR is around 2-3 ms, just like most temporal upscalers.

The number that Cerny said about 45% more performance is probably a best case scenario.
Most often than not, the Pro will be limited by memory bandwidth.

 

[zeroluck]

Nah. ive done several tests on this in several different games and it varies from game to game. some games cost 20% while others can go up as high as 35%.

I don't know what you mean by % cost.

 

[SweetTooth]

I bet the performance cost of PSSR is around 2-3 ms, just like most temporal upscalers.

The number that Cerny said about 45% more performance is probably a best case scenario.
Most often than not, the Pro will be limited by memory bandwidth.

Winjer chimed in with his wisdom, those deceiving cucks at Sony knows nothing about hardware design!!

How come you haven't been hired? Cmon Microsoft! Snag this genius

Edit: actually there are multiple geniuses in this thread! What are you doing Microsoft?!

 

[Grildon Tundy]

Assuming 35% faster gpu and no cpu bottleneck, ps5 pro without pssr should be getting 111 * 1.35 = 150 or 6.66ms frametime, with pssr its getting 102 or 9.8ms, so pssr takes 3.14ms, that is actually way slower than what I expected, a 3080 takes ~1ms to upscale from 1080p to 4k.

Is your 1ms upscaling DLSS? I wonder if the added time could be PSSR not being quite as optimized.

 

[DaleinCalgary]

 

[zeroluck]

Is your 1ms upscaling DLSS? I wonder if the added time could be PSSR not being quite as optimized.

Yeah, DLSS

 

[Haint]

This was pretty apparent when it was first revealed pssr games were largely running at the same res as base PS5, and even dynamic res wasn't really pushing render higher. A very very expensive solution, both in terms of processing and dollar cost ($800 console).

 

[squidilix]

Looking forward to this on pc. Il play it ultrawide 100+ fps and 4k like I did the first game.

 

[Zathalus]

He also compares TAA to PSSR. However, I'm not sure if TAA and PSSR are running the same input and output resolutions.

According to DF they are. In the next vid.

 

[Gaiff]

I bet the performance cost of PSSR is around 2-3 ms, just like most temporal upscalers.

The number that Cerny said about 45% more performance is probably a best case scenario.
Most often than not, the Pro will be limited by memory bandwidth.

Document said ~2ms to upscale from 1080p>4K.

 

[Three]

I don't know what you mean by % cost.

He means the the 35% differs from game to game so doing a theoretical 1.35 to figure out the delta would not be true and is an assumption. It could be 20% for this game making the delta smaller in your calculation.

 

[Embearded]

Wait, didn't they say PSSR runs on dedicated silicon?

 

[OverHeat]

Winjer chimed in with his wisdom, those deceiving cucks at Sony knows nothing about hardware design!!

How come you haven't been hired? Cmon Microsoft! Snag this genius

Edit: actually there are multiple geniuses in this thread! What are you doing Microsoft?!

WTF.gif seriously what is your problem? Your gonna get mass reported with this attitude.

 

[Senua]

WTF.gif seriously what is your problem?

The most hardcore Sony shill working today! He's an inspiration to us all

 

[Three]

Wait, didn't they say PSSR runs on dedicated silicon?

It has nothing to do with what silicon. It's added processing time before displaying the image thus it has a time in ms which affects framerate.

 

[vkbest]

Wait, didn't they say PSSR runs on dedicated silicon?

Yes, but GPU need wait to the upscaled frame. Unless they make their game full asynchronous, but I really doubt most developers would do

 

[Embearded]

It has nothing to do with what silicon. It's added processing time before displaying the image thus it has a time in ms which affects framerate.

Yes, but GPU need wait to the upscaled frame. Unless they make their game full asynchronous, but I really doubt most developers would do

Thanks, I understand that, it takes time in the pipeline of the frame. Some of the wording here confused me and i thought they meant it's an algorithm solution like FSR.

 

[Three]

Yes, but GPU need wait to the upscaled frame. Unless they make their game full asynchronous, but I really doubt most developers would do

How would you even do this with AI upscaling. Is it possible somehow?

 

[Seomel]

DF spoke about thid in new direct weekly and they have much better game to judge this by - GOW Ragnarok patch whoch dropped. You can select taa or pssr with unlocked fps, so they were able to get idea of cost of pssr
tldr- around 2ms latency cost, pssr resolves much better image. There isnfps difference but a lot of paces you are like 70-100range and with vrr it doesnt really have big difference in feel

 

[zeroluck]

How would you even do this with AI upscaling. Is it possible somehow?

Yes, ID tech did it with Wolfenstein Youngblood using async DLSS, DLSS only uses tensor cores, leaving shaders cores idle, if you have other tasks that don't use too much memory bandwidth you can run them on top of DLSS. If PSSR uses something akin to tensor cores then it should also be possible, if not then no.

 

[Three]

Yes, ID tech did it with Wolfenstein Youngblood using async DLSS, DLSS only uses tensor cores, leaving shaders cores idle, if you have other tasks that don't use too much memory bandwidth you can run them on top of DLSS. If PSSR uses something akin to tensor cores then it should also be possible, if not then no.

I'm not familiar with ID techs method. Doesn't the frame need to be ready before the upscale, how do you do it async?

 

[zeroluck]

I'm not familiar with ID techs method. Doesn't the frame need to be ready before the upscale, how do you do it async?

ID tech does presentation of frames in compute queue asynchronously, new frames don't need previous frame to finish before rendering. They overlapped upscaling with the rendering of the starting portion of next frame.

 

[SpokkX]

Yeah, DLSS

Nvidia is truly amazing. I use DLSS in every supported game and it always looks better but more importantly runs WAY better. Looking at that chart DLSS cost is almost not worth mentioning anymore - it is almost free. Imagine on 50-series

 

[SpokkX]

DF spoke about thid in new direct weekly and they have much better game to judge this by - GOW Ragnarok patch whoch dropped. You can select taa or pssr with unlocked fps, so they were able to get idea of cost of pssr
tldr- around 2ms latency cost, pssr resolves much better image. There isnfps difference but a lot of paces you are like 70-100range and with vrr it doesnt really have big difference in feel

So what you get lower fps with PSSR? Is it scaling from the same res as without?

 

[TrebleShot]

The Performance Pro mode on TLOU is the original Fidelity mode at a lower res then upscaled, so its not the original 1440p mode then upscaled.
It also has better shadows, lighting, models etc.

 

[SlimySnake]

The Performance Pro mode on TLOU is the original Fidelity mode at a lower res then upscaled, so its not the original 1440p mode then upscaled.
It also has better shadows, lighting, models etc.

Nope. Performance and fidelity modes were identical.

 

[Lysandros]

What happens when other optimized native applications even exceeds the stated 45% performance differential, words would be retracted? There are some which are afflicted by the addiction to assume the worst when it comes to Playstation hardware. Don't be consumed by the eagerness to jump into conclusions. Give the system some damned time to prove its capabilities.

 

[mrcroket]

It's funny how some people are in total denial about the fact that PS5 Pro is not as ‘Pro’ as expected, and is just a PS5 with a bit more raw power (which goes away when using PSSR) and AI rescaling for only $699 (disc drive and stand not included).

 

[Ironman_Bad]

Looking forward to this on pc. Il play it ultrawide 100+ fps and 4k like I did the first game.

Tell the waitress I'll take one of those medium well but hold the stutters please

 

[Ironman_Bad]

It's funny how some people are in total denial about the fact that PS5 Pro is not as ‘Pro’ as expected, and is just a PS5 with a bit more raw power (which goes away when using PSSR) and AI rescaling for only $699 (disc drive and stand not included).

Curious where you pulled that knowledge from?? Someplace that stinks I think

 

[OverHeat]

The most hardcore Sony shill working today! He's an inspiration to us all

Report is ass!

 

[kevboard]

this is still AMD hardware for the most part, so this wouldn't be a bit a surprise if it turns out to be accurate.

also interesting to note is that apparently in the code of the UE5 plugin for PSSR, PSSR uses nearly the exact same names and terminology for its variables and settings as those that previously were only used by XeSS.

which makes me wonder if PSSR is a modified version of XeSS that simply runs on the shaders. the base PS5 wouldn't be able to support XeSS at decent performance levels as it doesn't have hardware acceleration for DP4a, so an upgrade to a newer version of RDNA was necessary to make it work.
and the additional Compute Units were needed to mitigate the performance impact the DP4a path of XeSS still has when using big upscaling factors like a 4x scale from 1080p to 4k (since the more pixels that need to be created the higher the cost)

 

[Lysandros]

I just heard from a good friend who works at Sony PlayStation hardware engineering department that (rather bafflingly) employees there use two completely unique terminologies to refer to GPU shaders. One of them is "custom hardware for machine learning." And the other "An actual custom silicon AI upscaler." Well, that would explain everything.

 

[PicoLordens]

One thing I noticed while playing earlier is that on Pro mode the LOD draw distance is way worse than on fidelity mode. Very easily observable.

 

[Mr Moose]

this is still AMD hardware for the most part, so this wouldn't be a bit a surprise if it turns out to be accurate.

also interesting to note is that apparently in the code of the UE5 plugin for PSSR, PSSR uses nearly the exact same names and terminology for its variables and settings as those that previously were only used by XeSS.

which makes me wonder if PSSR is a modified version of XeSS that simply runs on the shaders. the base PS5 wouldn't be able to support XeSS at decent performance levels as it doesn't have hardware acceleration for DP4a, so an upgrade to a newer version of RDNA was necessary to make it work.
and the additional Compute Units were needed to mitigate the performance impact the DP4a path of XeSS still has when using big upscaling factors like a 4x scale from 1080p to 4k (since the more pixels that need to be created the higher the cost)

So because it uses similar naming in UE5, it must be XeSS?

 

[kevboard]

So because it uses similar naming in UE5, it must be XeSS?

it is a bit strange that all other upsampling methods have very distinct naming conventions. this is true for XeSS, DLSS, FSR, and TSR. PSSR is the only one that is this similar to another one, which happens to be XeSS.

so it is curious. and it wouldn't be that far fetched, as Intel already plans to eventually make it entirely open source.
so it wouldn't be surprising if Sony worked with Intel to fork off of it. especially since XeSS is very close to DLSS in quality in many games

 

[Mr Moose]

it is a bit strange that all other upsampling methods have very distinct naming conventions. this is true for XeSS, DLSS, FSR, and TSR. PSSR is the only one that is this similar to another one, which happens to be XeSS.

so it is curious. and it wouldn't be that far fetched, as Intel already plans to eventually make it entirely open source.
so it wouldn't be surprising if Sony worked with Intel to fork off of it. especially since XeSS is very close to DLSS in quality in many games

Oh yeah, I forgot they were using an Intel GPU.

 

[Chuck Berry]

Just saw this gaming tech video. Pros performance mode runs at 1440p internal resolution just like base PS5s performance mode. PSSR then reconstructs the image to 4k. Looks way better than 1440p.

However, the framerate comparison could help us determine just how much the PSSR reconstruction could cost on the GPU. The framerates are roughly in line on both the Pro and PS5 suggesting most of the circa 45% extra horsepower is going into PSSR reconstruction. In other Pro games, we have seen an extra 30% raw gpu performance so it could also be just 30%. In some cases, it dips around 10 fps below the ps5 performance mode like screenshotted below but that could also be a vram bottleneck since the textures and other effects are still rendered at 4k.

Timestamped:

Regardless, the reconstruction costs are in line with what we've seen from DLSS and FSR.

What’s your point Vanessa?