Support NeoGAF

[Mr.Phoenix]

Some of you need to stop low balling GPU compute so much. The PS5 Pro will follow the same route as the PS4 Pro, it's pretty simple.

They need to make the PS5 Pro seem worth it both in regards to on paper specs and in action. Trying to sell people on a 16 Tflop PS5 Pro without a disc drive for $599 ain't gonna cut it.

My pure guess would be 62-64 total CU with 60 active.

Also, where is this leak of 54 active CU? The leak was 60 CU, i.e. 30WGP.

Spec leaks never mention the inactive compute units. Look at leaks for OG PS4 and PS4 Pro before released, only active CUs were talked about.

So if the leaks hold true of 30WGP then that means PS5 Pro will have 60 active CUs.

I agree that tech leaks never mention inactive CUs. However, this leak also talked about 3 shader arrays of 10 work groups each. Th PS5 currently has two shader arrays. RDNA in the case of the PS5 (and XS) disables one workgroup in each shader array to improve yields. So if 3 shader arrays, and one WG disabled in each, we end up with 54CU.

But hey... even if we have 60CU active... let's be real here... that isn't doing what the PS4pro did. The PS4pro not only doubled the number of CUs from the OG to the pro, but it also increased clocks. For the PS5pro to do the same, it would have to have 72CU. All we are hearing is 60. And there is hardly a GPU on the market with a base clock of over 2400Mhz. And we know it cant be clocked at anything below the current PS5 GPU clocks. So that suggests that the PS5pro is clocked at best at 2300- 2400mhz, so let's just say 2350mhz.

So,
60CU @ 2350mhz = 18TF
54CU @ 2350mhz = 16.2TF

That is still a far cry from the 25TfF of the 7900XTX. And that's before RDNA3 compute magic math making it 50TF. But hey, if it's using the same RDNA3+ tech, then you can double that TF number and get 36TF and 32.4TF respectively. But I don't trust the whole RDNA3 VOPD thing, as its use case is generally very specific and limited thus far.

 

[SegaSnatcher]

I agree that tech leaks never mention inactive CUs. However, this leak also talked about 3 shader arrays of 10 work groups each. Th PS5 currently has two shader arrays. RDNA in the case of the PS5 (and XS) disables one workgroup in each shader array to improve yields. So if 3 shader arrays, and one WG disabled in each, we end up with 54CU.

But hey... even if we have 60CU active... let's be real here... that isn't doing what the PS4pro did. The PS4pro not only doubled the number of CUs from the OG to the pro, but it also increased clocks. For the PS5pro to do the same, it would have to have 72CU. All we are hearing is 60. And there is hardly a GPU on the market with a base clock of over 2400Mhz. And we know it cant be clocked at anything below the current PS5 GPU clocks. So that suggests that the PS5pro is clocked at best at 2300- 2400mhz, so let's just say 2350mhz.

So,
60CU @ 2350mhz = 18TF
54CU @ 2350mhz = 16.2TF

That is still a far cry from the 25TfF of the 7900XTX. And that's before RDNA3 compute magic math making it 50TF. But hey, if it's using the same RDNA3+ tech, then you can double that TF number and get 36TF and 32.4TF respectively. But I don't trust the whole RDNA3 VOPD thing, as its use case is generally very specific and limited thus far.

Who was expecting 25TF? All it needs to hit is around 20TF. That is totally doable for a console coming out in 2024 on likely a 4nm process.

Sony isn't dumb enough to try and sell people on a PS5 Pro that can't claim 2x the power.

 

[Panajev2001a]

Why should we wait when we can have it now (next year...)?

Because I do not think you can have it next year (the kicker with Pro consoles is that they are never and never become the target console so they are never truly optimised for).

It doesn't make any sense other than to satisfy some random imperative not to have mid gen consoles.

It is not random… We are seeing diminishing returns (performance improvements in HW vs IQ and resolution and framerate improvements are non linear), HW evolution is slowing down (it takes longer and longer to achieve each major manufacturing node jump and it is more and more expensive to design and manufacture chips on them), users expectations in terms of console pricing and size remain similar enough over the years.

Everything seems to suggest releasing more HW more frequently is not the answer, but the opposite would be.

As for the second point: it seems that we will (and to some extent have got) truly next gen graphics at something like 1080-1440p at 30fps. A Pro should be able to get close to that at 60, especially with VRR. But yes, that is essentially just a claim, and I don't know for sure. But it's a reasonable claim, prima facie.

Even this revised target (not sure how unstable 60 FPS @1080p saved only by VRR would make people that happy to have forked $499 for a PS5 Pro and ready to fork $499 again in a few years for PS6…) could be quite difficult to achieve because devs tend to only give token support to these machines so whatever you can get by bruteforcing through problems sure, but that just means you need an even bigger HW performance improvement delta to achieve things so it is more difficult not easier.

The games which are CPU limited doubling in framerate on average will also be interesting as Pro consoles tend to do very small changes to the CPU side of things to make BC easier (PS4 Pro just bought you what a 500 MHz CPU clocks bump? Significant in relative terms as the baseline was 1.6 GHz but it was not a magic cure… I will be surprised if PS5 Pro were to have a greater relative jump there). GPU wise let’s see, there is some chance there, but that is totally predicated on something beyond RDNA 3.5 as the next big change RT wise is rumoured to be with RDNA4.

 

[Thick Thighs Save Lives]

Wait, is this the "Slim" that's rumored to arrive later this year?

 

[Daniel_Mallorca1985]

Wait, is this the "Slim" that's rumored to arrive later this year?

Yes

 

[Mr.Phoenix]

Who was expecting 25TF? All it needs to hit is around 20TF. That is totally doable for a console coming out in 2024 on likely a 4nm process.

Sony isn't dumb enough to try and sell people on a PS5 Pro that can't claim 2x the power.

Ok cool...well its not hitting 20TF though. There simply is nothing so far to suggest that it will.

Only way it hits or exceeds 20TF, is if we are calculating those TFs using RDNA3 VOPD calculations. And if you know how that works you know how dubious those calculations are or have up to 72CU (which so far nothing says it will) or are clocking that 60CU at over 2600mhz (which sofar no GPU on the market even with their massive heat sinks clocks as high).

Oh they can claim 2x the power. It just won't be real. And sony doesn't have to sell people 2x the power. That's just some fanboy bragging rights-type nonsense. The PS5pro doesn't need to have 2x the TFnumber to yield overall better performance.

 

[CanYouFlyBobby?]

Ill ask again, how many teraflops are we talking here?

 

[SegaSnatcher]

Ok cool...well its not hitting 20TF though. There simply is nothing so far to suggest that it will.

Only way it hits or exceeds 20TF, is if we are calculating those TFs using RDNA3 VOPD calculations. And if you know how that works you know how dubious those calculations are or have up to 72CU (which so far nothing says it will) or are clocking that 60CU at over 2600mhz (which sofar no GPU on the market even with their massive heat sinks clocks as high).

Oh they can claim 2x the power. It just won't be real. And sony doesn't have to sell people 2x the power. That's just some fanboy bragging rights-type nonsense. The PS5pro doesn't need to have 2x the TFnumber to yield overall better performance.

Bare minimum it will be 18TF, there is no world where Sony pushes out a 15 - 16 TFlop "Pro" console.

Sure, it doesn't have to have 2x the power on paper to offer an improvement, but it's gonna be a harder sell when you can't claim it with numbers.

The market a Pro console is targeting are not ignorant about tech.

Again, nobody is paying $599 for a 15 Tflop Pro, we aren't in 2006 Sony era dumbness anymore.

At that point they'd be better off waiting until releasing the PS6.

 

[[Sigma]]

What is this abomination? First you reduce my copper and now your telling me no liquid metal? No thank you. I don't care that it runs cooler, put it in anyway!!

 

[bitbydeath]

 

[Robb]

Bring it on!

 

[mrBarrelNut]

IIRC RDNA3 had some 50% performance per watt improvement over RDNA2. I.e. it can pack 50% more performance in the same thermal envelope than RDNA2 based chip.

 

[lh032]

NO LIQUID METAL!!??
CERNY LYING ON THE COOLING VARIABLE CPU TECH

 

[ergem]

RDNA4 cards are supposedly releasing late next year. What are the odds AMD and Sony are developing a bespoke version for PS5 Pro along with the PC gpu counterparts?

 

[SmokSmog]

NO LIQUID METAL!!??
CERNY LYING ON THE COOLING VARIABLE CPU TECH

Liquid metal not needed in a 120~W console.

 

[StealthGoblin]

Given that the plates they are selling will likely fit this too, I don’t think this will be smaller in size.

 

[SmokSmog]

PS5 Pro
8 core Zen4c with 16MB unified L3 ( making it more cores than base ps5 is waste of die area, games will be designed for the base ps5)
RDNA3.5 AKA the fixed RDNA3 on TSMC N4 60CU GPU with 56CU active at 2.8ghz = 20TF FP32
256bit bus 16GB Vram 18Gb/s speed = 576GB/s memory bandwidth + 64MB L3 infinity cache. ???

BTW 7900XTX 96CU at 2500mhz is a 30TF FP32 GPU, dual issue doesn't even work on RDNA3.

 

[Physiognomonics]

With a monolithic design, it may be better the L3 cache doubles as Infinity Cache.

With this, I was aiming for around 300mm² on 5nm.
(For how I do these images, I insert the die shot image into a program name Autocad. Scale the image to the actual mm², then trace out the important units.)

~298mm2
12 Cores = 24MB L3 cache
Extra cores to boost fps at lower res.

18GB = 648 GB/s
8k may require bigger res/size textures.

Plus XDNA, since it reliefs the CPU of AI tasks.

I see you are thinking about at 5nm node for PS5 Pro. but if they are releasing a PS5 slim at 5nm this year, can't they use smaller node (like 4nm) next year? And what about Infinity cache? Shouldn't it be a big problem with BC?

 

[Hunnybun]

Because I do not think you can have it next year (the kicker with Pro consoles is that they are never and never become the target console so they are never truly optimised for).


It is not random… We are seeing diminishing returns (performance improvements in HW vs IQ and resolution and framerate improvements are non linear), HW evolution is slowing down (it takes longer and longer to achieve each major manufacturing node jump and it is more and more expensive to design and manufacture chips on them), users expectations in terms of console pricing and size remain similar enough over the years.

Everything seems to suggest releasing more HW more frequently is not the answer, but the opposite would be.


Even this revised target (not sure how unstable 60 FPS @1080p saved only by VRR would make people that happy to have forked $499 for a PS5 Pro and ready to fork $499 again in a few years for PS6…) could be quite difficult to achieve because devs tend to only give token support to these machines so whatever you can get by bruteforcing through problems sure, but that just means you need an even bigger HW performance improvement delta to achieve things so it is more difficult not easier.

The games which are CPU limited doubling in framerate on average will also be interesting as Pro consoles tend to do very small changes to the CPU side of things to make BC easier (PS4 Pro just bought you what a 500 MHz CPU clocks bump? Significant in relative terms as the baseline was 1.6 GHz but it was not a magic cure… I will be surprised if PS5 Pro were to have a greater relative jump there). GPU wise let’s see, there is some chance there, but that is totally predicated on something beyond RDNA 3.5 as the next big change RT wise is rumoured to be with RDNA4.

Well it basically comes down to you not thinking a 1.8x or so advance in GPU performance is possible, then.

If it's not, and performance modes are still like 900p, then yeah a Pro console probably wouldn't be worth the upgrade. If we can get more like 1080p-1440p, then imo it would.

Guess we'll see.

 

[Rocco Schiavone]

No liquid metal on the slim? Enjoy your dry shitty paste after one or two years.

But only if applied correctly.

 

[Shane89]

Pffff. I mean, it's good, more more power etc, and f**k me because i'll surely buy it, but that middle-gen consoles "trend" are really annoying me.

 

[SegaSnatcher]

No liquid metal on the slim? Enjoy your dry shitty paste after one or two years.

But only if applied correctly.

Where is the no liquid metal on slim rumor?

 

[Loxus]

You are increasing the Cpu Cores size by a lot btw, a Zen4C core+L2 uses 2.48mm of area. 8 of them uses 19,48mm on 5nm.

What do you mean?
CPUs are getting bigger. Zen4c actually stopped that.

PS5 Zen 2 Core Area: 2.50mm² @ 7nm
Zen 2 Core Area: 2.84mm² @ 7nm
Zen 3 Core Area: 4.05mm² @ 7nm
Zen 4 Core Area: 3.84mm² @ 5nm
Zen 4C Core Area: 2.48mm² @ 5nm

PS5 entire CPU area = ~38mm² @7nm
8 Zen4c CPU area = ~29mm² @5nm
12 Zen4c CPU area = ~44mm² @5nm
(CPU area = Core + L2 + L3)
Zen4c area is without cutting down the L3 cache.

PS5 entire GPU area = ~151mm² @7nm
60CU RDNA3 area = ~122mm² @5nm
(GPU area = Shader Engines + Command Processor + L2)

Take a look at Zen4c vs Zen4

Now imagine if Sony used Zen4 instead of Zen4c.

 

[SkylineRKR]

I don't see the use. Okay, it might do some RT at 1440p here and there instead of 1080p. Or 1080p/60fps instead of 30. But its not like 2016 when the TV landscape massively pushed 4K sets, the Pro made sense in a way.

 

[Loxus]

I understand that you are referring to the “”NO fat””

I don't know why Sony would ditch Liquid Metal.
It's so much better than thermal paste in every way.

Unless cost is becoming an issue.

 

[Loxus]

I see you are thinking about at 5nm node for PS5 Pro. but if they are releasing a PS5 slim at 5nm this year, can't they use smaller node (like 4nm) next year? And what about Infinity cache? Shouldn't it be a big problem with BC?

If you want 4nm, just reduce die area by around 10%.

Cache shouldn't be a problem.
Only thing that i know interferes with BC are GPU CUs and clock speeds.

 

[Mr.Phoenix]

Given that the plates they are selling will likely fit this too, I don’t think this will be smaller in size.

Or they could just make new plates for the PS5pro? The same way they would make new plates for the PS5sim...

PS5 Pro
8 core Zen4c with 16MB unified L3 ( making it more cores than base ps5 is waste of die area, games will be designed for the base ps5)
RDNA3.5 AKA the fixed RDNA3 on TSMC N4 60CU GPU with 56CU active at 2.8ghz = 20TF FP32
256bit bus 16GB Vram 18Gb/s speed = 576GB/s memory bandwidth + 64MB L3 infinity cache. ???

BTW 7900XTX 96CU at 2500mhz is a 30TF FP32 GPU, dual issue doesn't even work on RDNA3.

This 4nm thing again. And 2.8TF? Nope.

I see you are thinking about at 5nm node for PS5 Pro. but if they are releasing a PS5 slim at 5nm this year, can't they use smaller node (like 4nm) next year? And what about Infinity cache? Shouldn't it be a big problem with BC?

The newest and smallest node is usually the most expensive one too. As or an infinity cache if they do one, its a simple case of replacing one of the cache levels with it. So where in the current PS5 you have an l2 cache of say...4MB, on the pro it would be 32MB/64MB. It would be transparent to the games and its use would be on a system level. Kinda like having a faster drive in a computer.

It just means the PSoS/scheduler/profiler or whatever is responsible for it would have a less aggressive cache scrubbing routine on the PS5pro, at the end of the day, infinity cache boosts memory bandwidth simply by reducing how often you have to go all the way to main RAM to pull data or by increasing your cache hit rate.

I don't know why Sony would ditch Liquid Metal.
It's so much better than thermal paste in every way.

Unless cost is becoming an issue.

It's also more complicated to work with, and you really don't need the most advanced thermal interface to cool a sub-150w APU.

 

[Physiognomonics]

Or they could just make new plates for the PS5pro? The same way they would make new plates for the PS5sim...

This 4nm thing again. And 2.8TF? Nope.

The newest and smallest node is usually the most expensive one too. As or an infinity cache if they do one, its a simple case of replacing one of the cache levels with it. So where in the current PS5 you have an l2 cache of say...4MB, on the pro it would be 32MB/64MB. It would be transparent to the games and its use would be on a system level. Kinda like having a faster drive in a computer.

It just means the PSoS/scheduler/profiler or whatever is responsible for it would have a less aggressive cache scrubbing routine on the PS5pro, at the end of the day, infinity cache boosts memory bandwidth simply by reducing how often you have to go all the way to main RAM to pull data or by increasing your cache hit rate.


It's also more complicated to work with, and you really don't need the most advanced thermal interface to cool a sub-150w APU.

This is why I hope there will be some L3 cache on the GPU. Otherwise this machine will have serious trouble with bandwith with only a 29% improvement compared to PS5 notably with the rumored new RT units.

 

[SmokSmog]

N4 is just a N5+++, by end of 2024 Nvidia will be moving to a proper node called N3E.

 

[ToTTenTranz]

I did 2 mockups of the PS5 Pro using RDNA3 and Zen4c.

One is 8 Zen4c cores and the other is 16 Zen4c cores.

5nm GCD - 6nm MCDs
8 core - 244mm²
16 core - 291mm²
MCD - 37mm² for each MCD.

These are the best case to utilize die space.

I'll try to do a monolithic design went I get the time.

Nice work.
I wouldn't assume another 128bit of GDDR6 just for 8 CPU cores, especially if the CPU can access the Infinity Cache.

 

[XesqueVara]

What do you mean?
CPUs are getting bigger. Zen4c actually stopped that.

PS5 Zen 2 Core Area: 2.50mm² @ 7nm
Zen 2 Core Area: 2.84mm² @ 7nm
Zen 3 Core Area: 4.05mm² @ 7nm
Zen 4 Core Area: 3.84mm² @ 5nm
Zen 4C Core Area: 2.48mm² @ 5nm

PS5 entire CPU area = ~38mm² @7nm
8 Zen4c CPU area = ~29mm² @5nm
12 Zen4c CPU area = ~44mm² @5nm
(CPU area = Core + L2 + L3)
Zen4c area is without cutting down the L3 cache.

PS5 entire GPU area = ~151mm² @7nm
60CU RDNA3 area = ~122mm² @5nm
(GPU area = Shader Engines + Command Processor + L2)

Take a look at Zen

How did you got 244mm for 8 cores but 291mm more for 16 ones?

 

[PeteBull]

I don't know why Sony would ditch Liquid Metal.
It's so much better than thermal paste in every way.

Unless cost is becoming an issue.

if its 5nm and doesnt eat 220W of power but 120-140W then its probable coz of the cost, yeh.

 

[Loxus]

How did you got 244mm for 8 cores but 291mm more for 16 ones?

44mm² is just for CPU Cores.
291mm² is the entire SOC.

Also, I added more memory controllers.

 

[Ainem Enamas]

It's simply an enhancer for games.

 

[Panajev2001a]

Well it basically comes down to you not thinking a 1.8x or so advance in GPU performance is possible, then.

I am saying that a 1.8x performance improvement in the GPU (specs wise) is possible perhaps (smaller than the jump from PS4 to to PS4 Pro btw and that should tell you something too), but that the relationship with game’s performance and IQ is not linear and getting worse.

 

[TheRedRiders]

Keplar weighing in on the recent leaks.

 

[Dokku]

What does this mean in English?

 

[squidilix]

What does this mean in English?

Knack 3 going to kick ass

 

[TLZ]

What does this mean in English?

Don't listen to others and only to Tom Henderson, basically.

Who's Kepler anyway?

 

[Investor9872]

Bare minimum it will be 18TF, there is no world where Sony pushes out a 15 - 16 TFlop "Pro" console.

Again, nobody is paying $599 for a 15 Tflop Pro, we aren't in 2006 Sony era dumbness anymore.

At that point they'd be better off waiting until releasing the PS6.

Like hell we ain't! We're living in the era where Sony is charging $300+ for earbuds, and $200 for a PS5 controller. As long as people keep buying at these prices, I wouldn't put anything pass Sony.

 

[SlimySnake]

Don't listen to others and only to Tom Henderson, basically.

Who's Kepler anyway?

But tom was the one who said 60 CUs.

 

[Investor9872]

Liquid metal not needed in a 120~W console.

I don't care. After having purchase the PS4 Pro on day-1, and having to put up with jet engine noise all these years, I want that liquid metal in there regardless.

 

[DenchDeckard]

I agree that tech leaks never mention inactive CUs. However, this leak also talked about 3 shader arrays of 10 work groups each. Th PS5 currently has two shader arrays. RDNA in the case of the PS5 (and XS) disables one workgroup in each shader array to improve yields. So if 3 shader arrays, and one WG disabled in each, we end up with 54CU.

But hey... even if we have 60CU active... let's be real here... that isn't doing what the PS4pro did. The PS4pro not only doubled the number of CUs from the OG to the pro, but it also increased clocks. For the PS5pro to do the same, it would have to have 72CU. All we are hearing is 60. And there is hardly a GPU on the market with a base clock of over 2400Mhz. And we know it cant be clocked at anything below the current PS5 GPU clocks. So that suggests that the PS5pro is clocked at best at 2300- 2400mhz, so let's just say 2350mhz.

So,
60CU @ 2350mhz = 18TF
54CU @ 2350mhz = 16.2TF

That is still a far cry from the 25TfF of the 7900XTX. And that's before RDNA3 compute magic math making it 50TF. But hey, if it's using the same RDNA3+ tech, then you can double that TF number and get 36TF and 32.4TF respectively. But I don't trust the whole RDNA3 VOPD thing, as its use case is generally very specific and limited thus far.

Who was expecting 25TF? All it needs to hit is around 20TF. That is totally doable for a console coming out in 2024 on likely a 4nm process.

Sony isn't dumb enough to try and sell people on a PS5 Pro that can't claim 2x the power.

I think 18tf would be great for a ps5 Pro if it hits like 599 dollars. Now if it was my dream of 799 then I would expect something like 25 tf.

Also, just to add...wasn't the ps4 pro a special part where they made the butterfly effect and doubled up the CUs.

I think that Sony could convince AMD to do the same again and offer a cheap 72 CU design for them where they just butterfly wing two 36 CUs?

I know nothing and I'm an idiot with tech stuff on this kind of level. Just throwing ideas out there.

I'd happily pay 599 for an 18tf ps5 Pro. Quite excited for it tbh.

 

[TheRedRiders]

Don't listen to others and only to Tom Henderson, basically.

Who's Kepler anyway?

Reliable AMD leaker. He and Tom were the only leakers who were aware of the PS5 Pro's codename prior to Tom's official leak.

He doesn't have a YT channel either so you know he's not making shit up for clout.

 

[Insane Metal]

But tom was the one who said 60 CUs.

Wasn't it 60 WGPs? That's different from 60CUs.

 

[THE:MILKMAN]

Wasn't it 60 WGPs? That's different from 60CUs.

30 WGP's which equates to 60CU's. Curious what he actually means by this being a standard Sony design?

 

[Crispy Gamer]

Wasn't it 60 WGPs? That's different from 60CUs.

30 WGP and everyone assumed 60 CU but it's supposedly on RDNA 3.5+

 

[Insane Metal]

30 WGP's which equates to 60CU's. Curious what he actually means by this being a standard Sony design?

Thanks for the correction.

 

[THE:MILKMAN]

30 WGP and everyone assumed 60 CU but it's supposedly on RDNA 3.5+

That wouldn't fit the "standard Sony design" though would it? PS4 Pro actually kept the old GCN arch and added Polaris tech didn't it? So wouldn't PS5 Pro follow with it being predominately RDNA 2-based with tech added from future architectures and Cerny sauce like stronger RT?

 

[midasmulligan]

Think it’ll be able to hit 900p?? DLSS to the moon baby, yeah!!!