• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
  • The Politics forum has been nuked. Please do not bring political discussion to the rest of the site, or you will be removed. Thanks.

Analysis Hardware Why did Sony stay with 36CUs for the PS5?

Xplainin

Banned
Apr 30, 2020
700
1,528
545
I think we were all surprised that Sony stuck with the same amount of compute units with the PS5 as they did with the PS4 Pro.
It has been said that it was to help with back compate for PS4 games, but I'm not 100% sold on that alone.
I'm not a dev so could there be other reasons such as making it easier to develop for by keeping the same amount of Cu's?
Like how much change would you have to do to dev tools by increasing the amount of Cu's?
Would keeping the same 36CUs mean their internal studios engines would be quicker to adapt to the PS5?

Or was it simply that like with the PS4 -> Pro that they would have needed to have either 36 Cu's on a butterfly GPU with 72CUs and as 72 was unfiesable it had to be 36?
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
Because the clocks are higher and that's more important.

so people still believe this PR crap, ok well... PC tests show they are not important, but hey... why look at empirical evidence?

@topic: well, it's cheapter, that's basically what it comes down to. it most likely also has to do with the fact that developing a console takes a very long time, meaning if you plan for more CUs and the technology isn't there once you want to launch the system, you're fucked.
Sony knew that RDNA can support up to 40CUs and 40CU chips have been made by AMD. so if they wanted to make sure their goals will be met by the time their planned release comes closer they can go the safe route and simply use 40CU chips to make extra sure that their system can release on time and that they can give their developers a set target hardware as fast as possible and as closely as possible to final silicon.

that was most likely why... a mix of cost reduction and early+safe planning
 
Last edited:
Mar 27, 2020
14,338
37,067
725
Uncharted bunghole
I think it was said that Cerny looked at 42 CUs as well. But he choose a lower CU count because he wanted the GPU to run at a faster speed. Technically you could clock 42 CUs higher but with the additional compute units you also get additional heat.

My opinion.

They toyed with multiple CU counts with a predetermined amount of space for the die. For them 32 CUs hit the sweet spot and got Cernys the high clocks that he wanted. I also believe they had to reserve some of the die space for the I/O complex which takes up a considerable amount of space. They wanted to have that as well which less less space for CUs.
 
Last edited:

Thicc Enigma

Member
Oct 21, 2013
641
1,525
920
From what I recall and someone correct me if I'm wrong but I think he said they needed to stay at 36CU in order to be able to have BC with PS4 (PS4 Pro CU count is also at 36 while base PS4 CU count is at half that at 18) Something about running at same frequencies. Not sure.
 
  • LOL
  • Like
Reactions: jonnyp and ZehDon

S0ULZB0URNE

Member
Jan 27, 2020
4,424
5,654
660
so people still believe this PR crap, ok well... PC tests show they are not important, but hey... why look at empirical evidence?



that was most likely why... a mix of cost reduction and early+safe planning
Thats not true.
The PS5 is more likely to be maxed out early on with the set up it has.

Hmm further elaboration from you 2 be nice.

I need a good chuckle.
 

Craig of War

Member
Apr 19, 2020
2,270
7,749
735
Antarctica
From what I recall and someone correct me if I'm wrong but I think he said they needed to stay at 36CU in order to be able to have BC with PS4 (PS4 Pro CU count is also at 36 while base PS4 CU count is at half that at 18) Something about running at same frequencies. Not sure.
Really that just weird how come MS does not do that and still have better Backwards Compatibility.
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
Hmm further elaboration from you 2 be nice.

I need a good chuckle.

tests done with RDNA PC graphics cards, one running at lower clocks but higher CUs and one running lower CU count but higher clocks, show the performance difference is basically not measurable in almost any game, and is actually pointing towards more CUs having a positive effect on some.
so there is no evidence in this actually having any positive impact on modern game engines on RDNA PC parts
 
Last edited:
Oct 26, 2018
22,582
31,739
885
I remember lots of people saying Sony matched 36 CU to Pro since it makes it easier to do BC with PS4 games. Not sure how true that is.
 
  • LOL
Reactions: jonnyp

oldergamer

Member
Aug 20, 2004
4,084
4,082
1,605
They had no choice. Sony is locked into some choices and haven't developed proper hardware abstraction for the OS and development tools. Its going to affect the CU count going forward.

This is the reason they need a "ps4 mode" where you don't get a benefit of increased speed with the new hardware. Xbox doesn't have this problem based on the way the software is structured. This is why xbox is able to provide visual upgrades for older games that were not developed for newer hardware.
 

onQ123

Member
May 1, 2010
17,755
10,750
1,460
Probably for the same reason that AMD don't have any gaming GPU's over 40CU
 
Jun 10, 2020
276
386
340
Probably a mix of budget (less CUs means cheaper to produce and theoretically better yields), ease if backwards compatibility in the hardware (PS4 Pro had double PS4 and in BC mode shut off half the GPU, and matching CU counts for PS5 likely allows easier support for the virtualization team) and Mark Cerny’s personal preference of lower CU count clocked higher.


Nobody really knows, and if we ever find out, it won’t be for another 10 or so years
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
They had no choice. Sony is locked into some choices and haven't developed proper hardware abstraction for the OS and development tools. Its going to affect the CU count going forward.

This is the reason they need a "ps4 mode" where you don't get a benefit of increased speed with the new hardware. Xbox doesn't have this problem based on the way the software is structured. This is why xbox is able to provide visual upgrades for older games that were not developed for newer hardware.

that's not true, PS4 games will be able to use the full speed of the hardware. base PS4 games with no Pro patch will most likely only be able to use half of the CUs tho, but at full clock speeds. Sony said that they work towards most PS4 games being able to run in boost mode (high clocks)

the difference here is that Xbox One games will run using the full hardware of the Series X, not only the increased clock speeds.
 
  • Like
Reactions: PPIMystaceus

jakinov

Member
Nov 19, 2008
1,096
1,228
1,260
Mark Cerny explains their reasoning in the PS5 video.

1. PS4 CU is not the same as a PS5 CU, they are bigger, more efficient and FLOPS isn’t a good calculation with different architectures

2. other parts of the GPU run faster when you have a higher clock speed

3. when a developer has many CUs he argues that it’s harder to keep all your CUs busy with work with modern games.
 
Last edited:

psorcerer

Banned
May 1, 2012
2,479
3,537
1,035
tests done with RDNA PC graphics cards, one running at lower clocks but higher CUs and one running lower CU count but higher clocks, show the performance difference is basically not measurable in almost any game, and is actually pointing towards more CUs having a positive effect on some.
so there is no evidence in this actually having any positive impact on modern game engines on RDNA PC parts

You have no access to any CUs on PC.
What exactly was measured?
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
3. when a developer has many CUs he argues that it’s harder to keep all your CUs busy with work with modern games.

that is theoretically true, but once again, PC tests show that for modern engines that is simply not the case... they easily use all available cores of a GPU, which is why you see an RTX2080ti outperform an RTX2080 at same clock speeds. if games wouldn't use the power of the additional hardware the games would run the same, which they don't

You have no access to any CUs on PC.
What exactly was measured?

you run the same game on 2 GPUs, one running less active CUs at higher clocks and one running more CUs at lower clocks. both have the same peak TF performance. if high clocks would benefit the game engine the games would run better on the lower CU higher clock GPU, but they don't.
 
Last edited:

SlimySnake

Member
Feb 5, 2013
12,734
36,113
1,260
they were trying to target a cheaper price point. likely $399. they have no lockhart to fall back on. so they had to compromise where as microsoft didnt and aimed to make a $500 console from the beginning.
 

psorcerer

Banned
May 1, 2012
2,479
3,537
1,035
you run the same game on 2 GPUs, one running less active CUs at higher clocks and one running more CUs at lower clocks. both have the same peak TF performance. if high clocks would benefit the game engine the games would run better on the lower CU higher clock GPU, but they don't.

Again. It just measures how good DX/Vulkan abstract the hardware. Hint: pretty good.
 

Investor9872

Member
Jan 4, 2020
1,140
1,519
410
Thats not true.
The PS5 is more likely to be maxed out early on with the set up it has.

That's one way to put it. But I trust Mark Cerny's vision, and I'm sure he knows what's he doing. My guess for keeping with 36 CUs so that it would work more efficiently with all the other custom parts. Remember that the PS5 CUs are of larger design than the XB Series X's CUs.
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
Again. It just measures how good DX/Vulkan abstract the hardware. Hint: pretty good.

this is important, because if any developer could actually use the supposed benefits of higher clocks and lower CUs it would be first party games since third party studios usually use GNMX, which is a high level API that isn't capable of going down to the metal to tailor the engine directly to the hardware. while first party devs usually use GNM, which is their low level API.

this means that third party developers will also use an API that is comparable to Direct X, they will not really benefit from this. first aprty developers use the low level API and only work with a single piece of hardware, meaning that no matter how wide the GPU would be, they would be able to fully utilize it and the narrow but fast design really doesn't help much. their reason given is that a wide design makes it harder to fully utilize each CU, if their direct X esque API does that on its own tho and their first party devs have the ability to utilize whatever they get... this is not really an argument.

Remember that the PS5 CUs are of larger design than the XB Series X's CUs.

where the fuck did you get that from? that's simply not the case... I am really confused where this is coming from tho. first time hearing that tbh and not quite sure what it's supposed to mean either.
 
Last edited:

Dee_Dee

Member
Apr 22, 2020
384
833
265
Thats not true.
The PS5 is more likely to be maxed out early on with the set up it has.
Yea i can see that. Remember a lot devs said it’s like the easiest console to develop on. I expect to see some mind blowing looking games (especially exclusives)on PS5 In it’s first year.
I don’t think XSX exclusives will look as good or better than PS5 exclusives until they kill off the Xbox one. So probably 2023 unless a PS5 pro is released that year.
 

S0ULZB0URNE

Member
Jan 27, 2020
4,424
5,654
660
tests done with RDNA PC graphics cards, one running at lower clocks but higher CUs and one running lower CU count but higher clocks, show the performance difference is basically not measurable in almost any game, and is actually pointing towards more CUs having a positive effect on some.
so there is no evidence in this actually having any positive impact on modern game engines on RDNA PC parts
No RDNA 2 cards are available.
PS5's GPU/Chipset is a CUSTOM solution so it's impossible to use PC based parts to see how it would perform.

In my experience of my non PS5 PC parts... higher clocks gave me significant increases.
 
Last edited:

TearInto5th

Member
Mar 8, 2020
149
357
270
Because they dont need to go higher. The difference in Tflops will not make a this massive difference that people claim.

Theyve designed their system to go with the focuses of their software. Epic AAA single player masterpieces. Have put in the extra effort to go redesign everything in it to make that software strength shine.

I think the PS5 will have better looking games exclusives wise, because I dont think MS has the patience to put the effort in.
 

S0ULZB0URNE

Member
Jan 27, 2020
4,424
5,654
660
that's not true, PS4 games will be able to use the full speed of the hardware. base PS4 games with no Pro patch will most likely only be able to use half of the CUs tho, but at full clock speeds. Sony said that they work towards most PS4 games being able to run in boost mode (high clocks)

the difference here is that Xbox One games will run using the full hardware of the Series X, not only the increased clock speeds.
XSX is using SOFTWARE based EMULATION.
 
  • Like
Reactions: Devilich

01011001

Member
Dec 4, 2018
8,250
12,896
620
No RDNA 2 cards are available.
PS5's GPU/Chipset is a CUSTOM solution so it's impossible to use PC based parts to see how it would perform.

In my experience of my non PS5 PC parts... higher clocks gave me a significant increases.

1: RDNA 2 is not a paradigm shift from RDNA 1, it is an evolution of it. this means they are very comparable.
2: current gen consoles compare to their respective PC parts pretty much 1 to 1 even tho they are modified. only a slight performance increase due to console optimizations are noticeable really. and as engines become more and more optimised for cross platform devlelopment we will see a reduction of these optimization benefits
3: so you had 2 PCs, one of them using a narrower GPU and one using a wider GPU clocked at the exact speeds so both would have the same theoretical teraflop performance? I don't know why I doubt that you did...

XSX is using SOFTWARE based EMULATION.

it is... to a point, at least with 360 games. we don't know with Xbox One games, an I would expect them to simply run natively with the API dong most of the heavy lifting when it comes to running the code.
also the Xbox One had hardware support for 360 software to a degree, meaning this was actually planned before the system launched, at least to some degree. so it is not full Software emulation there either.

not sure why you needed to point it out either way tbh. in the end the end user shouldn't really care how it works just how the end result turns out. which we will have to wait and see on both.
 
Last edited:

EventHorizon

Member
Jul 15, 2013
3,712
423
645
so people still believe this PR crap, ok well... PC tests show they are not important, but hey... why look at empirical evidence?

Those tests are like testing games late in the PS3/360 generation and concluding that the 360's easier to develop for architecture didn't provide any benefit, when the reality was that devs had to jump through hoops to negate the PS3's deficiencies. The same thing is going on here. Using fewer CUs clocked faster does make development easier, and that will show benefits to the PS5, especially early in the generation.

Also note that PCs have their own bottlenecks that won't be applicable to next gen consoles. Your argument sounds like one I heard where the person claimed the PS5's faster SSD won't make any appreciable difference in games since tests with faster SSDs on PCs showed little difference. What that failed to take into account was that the PC I/O API was a bottleneck, which is why DirectStorage is now a thing, and next gen consoles won't have that bottleneck.

Likewise the PS5 API and development best practices will make full use of its faster GPU clocks, which increase all aspects of the GPU not just a subset. Games and game engines, like Unreal Engine 5, will do so also. All of this is especially important on consoles, which are standardized and coded closer to the metal. In short, current gen tests of PC games will have little relevance to the performance of next gen console games designed specifically to maximize next gen console hardware and running on APIs designed to facilitate it.
 
Last edited:

S0ULZB0URNE

Member
Jan 27, 2020
4,424
5,654
660
1: RDNA 2 is not a paradigm shift from RDNA 1, it is an evolution of it. this means they are very comparable.
2: current gen consoles compare to their respective PC parts pretty much 1 to 1 even tho they are modified. only a slight performance increase due to console optimizations are noticeable really.
3: so you had 2 PCs, one of them using a narrower GPU and one using a wider GPU clocked at the exact speeds so both would have the same theoretical teraflop performance? I don't know why I doubt that you did...
Nope RDNA 2.0 is faster with more features.
Custom is still custom therefore can't be tested on PC.

Older GPU's that I had clocked higher didn't net the results as good as my other newer cards.
But again PS5 is using CUSTOM components and can't be compared 1:1 with the PC.
 

MarcCppDev

Neo Member
Jun 2, 2020
36
17
100
From what I recall and someone correct me if I'm wrong but I think he said they needed to stay at 36CU in order to be able to have BC with PS4 (PS4 Pro CU count is also at 36 while base PS4 CU count is at half that at 18) Something about running at same frequencies. Not sure.
Yeah they also downclock the GPU to PS4 clock speeds for back compat. Its a poor's man method to implement back compat IMO. MSFT software emulation is better IMO
 

01011001

Member
Dec 4, 2018
8,250
12,896
620
Nope RDNA 2.0 is faster with more features.
Custom is still custom therefore can't be tested on PC.

Older GPU's that I had clocked higher didn't net the results as good as my other newer cards.
But again PS5 is using CUSTOM components and can't be compared 1:1 with the PC.

we can still look at the nearest comparable hardware on PC and look if there are any benefits to be fund here by going with a narrow and fast design... which we don't.
this means the only source we have for that to be the case with modern game engines is the word of a company that would really like you to believe that that's the case.
 
  • Like
Reactions: CyRuZ and smbu2000

S0ULZB0URNE

Member
Jan 27, 2020
4,424
5,654
660
it is... to a point, at least with 360 games. we don't know with Xbox One games, an I would expect them to simply run natively with the API dong most of the heavy lifting when it comes to running the code.
also the Xbox One had hardware support for 360 software to a degree, meaning this was actually planned before the system launched, at least to some degree. so it is not full Software emulation there either.

not sure why you needed to point it out either way tbh. in the end the end user shouldn't really care how it works just how the end result turns out. which we will have to wait and see on both.
You said otherwise is why.
MS has never said it was using hardware based BC.

You are correct it shouldn't matter how they achieve it but we must come correct.

MS has a EXCELLENT BC solution but this point is null in this convo.
 

Xplainin

Banned
Apr 30, 2020
700
1,528
545
Mark Cerny explained it in Road to PS5. The answer was for higher clocks as the entire hardware benefits from it.

‘Rising tide lifts all boats.’

Obviously more CU’s would raise heat so it was about finding the sweet spot.
Let's be honest, you are going to sell what you have. I take that with a grain of salt, and the tests have shown better performance with more CUs and lower clocks.

I am actually looking for other reasons. Like if they stick with the same amount of Cu's, is there a benefit to their tools and engines?
Are Sony engines built around their GPU CU number, and so keeping the same amount means less work on that?

Or was it totally BC?