Discussion Playstation 6 speculation

[MrMPFR]

But not really much over the PS5 Pro.

I think the GPU clock will be in the 2.5 Ghz range.

Yeh

If true Sony gonna do PS4 form factor. 2.5ghz on N3P will be crazy efficient.

 

[MrMPFR]

Post #21's PS6 extrapolation math assumes RB and Rasterizer related pipeline is unchanged. Now Kepler has confirmed TBIMR for GFX13. In addition it's possible we'll see the inclusion of additional features such as TBB, TLPBB and more clever cachemem management for rendering pipeline.
The perf scaling issues from 9070 -> 9070XT I mentioned previously are prob related to something else than RB + Rasterizer being unchanged (except for higher clocks), which makes it possible to address with PS6 despite -25% SEs (3 vs 4).
This obviously changes everything and reaching the 9070XT perf target seems possible even at lowered clocks and power budget.

I expect PS6 to be comparable to AT3 / Medusa Halo

You're right and I was wrong it seems. New pipeline allows AMD to do more with less + SE issue was overblown. -33% SE (also RB + Rasterizer) vs PS6 won't matter that much. Kepler also confirmed expanded dual issue for RDNA5, that'll help boost IPC a lot.
Now that AT4 dGPU should be ~9060XT according to Kepler the 2X bigger AT3 dGPU should indeed be comparable to the PS6. Not sure about Medusa Halo though, because then it'll have to be crazy efficient.
We'll see.

 

[ToTTenTranz]

Now that AT4 dGPU should be ~9060XT according to Kepler the 2X bigger AT3 dGPU should indeed be comparable to the PS6.

I'd say the 48CU AT3 would need to be ~12% higher clocked than the PS6's 54CU for them to be comparable.

Looking at how Sony bet a lot on clockspeeds for the PS5, I don't know if that will be the case.

 

[marees]

I'd say the 48CU AT3 would need to be ~12% higher clocked than the PS6's 54CU for them to be comparable.

Looking at how Sony bet a lot on clockspeeds for the PS5, I don't know if that will be the case.

the power budgets are vastly different between 48cu AT4 vs 54 CU PS6

One place the PS6 will have the edge is in using gddr7 memory. sony has always gone for bonkers memory. so in memory bound scenarios the PS6 should come ahead otherwise it will be in ballpark of AT4

 

[ToTTenTranz]

One place the PS6 will have the edge is in using gddr7 memory. sony has always gone for bonkers memory. so in memory bound scenarios the PS6 should come ahead

Actually, if the AT3 dGPU uses 16 channels * 24bit LPDDR6 14.4Gbps then its RAM bandwidth will be 691GB/s. If the PS6 goes with 160bit GDDR7 as currently rumored, then it's 640GB/s.

Things will change drastically if the AT3 dGPU only uses LPDDR5X, though.

 

[MrMPFR]

I'd say the 48CU AT3 would need to be ~12% higher clocked than the PS6's 54CU for them to be comparable.

Looking at how Sony bet a lot on clockspeeds for the PS5, I don't know if that will be the case.

Thought PS6 was 52CUs with 2 disabled.

Nah still nothing compared to PC.
PS5 max clk 2.23ghz
RX 6700 XT FE = avg 2.49ghz
(2.23-2.49)/2.23 * 100 = >11.66% higher clocks.

Sorry for the following math but I think it's pretty interesting:
If I use 16% IPC from 15% clockspeed bump from RDNA5 thread = 3.57ghz (9060XT ~3.1ghz)
TPU 4K native raster 9060XT -> 9070XT = 1.83x perf
RDNA5 suffers from poor intra-SE CU scaling & clock scaling (9070 -> 9070XT) + non perfect SE scaling (9060XT -> 9070XT should be ~1.92X (clock adjusted) not 1.83X).

RDNA5 prob has localized everything (except GCP) and scales perfectly from 1 -> 2 SE unlike RDNA4. So 1.83/2 * 3.57ghz = 3.27ghz
That's 2 SE x 12 CU/WGP = 9070XT perf. Pretty neat.

CU for AT3 -> PS6 = +8.33%
PS6 has an advantage in +50% non-shader logic + shorter SE (easier to feed) so let's add what appears to be another 2 CUs. +12.5%. Could be more but IDK.
3.27ghz/12.5% = 2.9ghz 52CU PS6 GPU at 9070XT perf

From prev math at iso-clocks 9070XT has +23% CU lead. with 2982mhz/2900mhz (clk difference) x 23% = +26.48% IPC vs 9070XT.

As you can see perfect SE scaling + removing other bottlenecks really pays off and if f Sony drops clocks a bit more on GPU and CPU then they can make PS4 sized quiet console (lower BOM).

 

[MrMPFR]

the power budgets are vastly different between 48cu AT4 vs 54 CU PS6

One place the PS6 will have the edge is in using gddr7 memory. sony has always gone for bonkers memory. so in memory bound scenarios the PS6 should come ahead otherwise it will be in ballpark of AT4

Yeah GDDR7 gonna guzzle electrons in comparison to 384bit LPDDR6 But -370mhz will bring power down that's for sure.

AT4 prob just gonna go larger L2 to counteract that. IIRC PS6 only has 10MB L2.
Now there's also TBIMR, universal compression, overhauled scheduling, tons of cachemem changes (efficiency mostly) and also maybe local everything (except GCP), clean slate cache (Flexcache + global sharing of everything + shared data cache lines), TLPBB and prob more.
All this provides absolutely ludicrous memory BW savings. Doubt any of them will have mem BW issues.

Actually, if the AT3 dGPU uses 16 channels * 24bit LPDDR6 14.4Gbps then its RAM bandwidth will be 691GB/s. If the PS6 goes with 160bit GDDR7 as currently rumored, then it's 640GB/s.

Things will change drastically if the AT3 dGPU only uses LPDDR5X, though.

LPDDR6 has ECC overhead. 256/288 x 640GB/s = 569GB/s. I don't expect it to tap out LPDDR6 standard though but perhaps things will change ~2-2.5 years from now. Maybe ~13gbps. ~500GB/s.

32gbps over 160bit GDDR7 sounds plausible with all the architectural changes in mind.

Yeah but that would be incredibly stupid. -33% bit PHY and slower pin speed xD

 

[marees]

ML upscaling in playstation 6 will be a huge leap forward

https://twitter.com/x/status/2017709965209723160

 

[MrMPFR]

ML upscaling in playstation 6 will be a huge leap forward

https://twitter.com/x/status/2017709965209723160

Could be. Nothing is confirmed it's just speculation based on AMD patents and research papers.

But if me theory is correct, that is GFX13 goes well beyond GFX12.5 cachemem, then PS6 and ATx lineup will annihilate 50 series in effective ML perf.

 

[MrMPFR]

MLID continuing the PS6 = RDNA 5 and PS5 = RDNA 2 claims and saying there's no such thing as full RDNA5 in latest Broken Silicon.
Claims everything is custom and uses 7600 as an example, a different situation to PS5, that actually lacked feature set, here only differentiating thing vs RDNA3 MCM is VGPR size. Comparison would only make sense if small RDNA3 lacked in actual feature set, e.g. no workgraphs support.

Given the 99% RDNA5 claim by @Kepler_L2 do any of you have an idea of what the PS6 could be missing that ATx lineup will have?

 

[ToTTenTranz]

MLID continuing the PS6 = RDNA 5 and PS5 = RDNA 2 claims and saying there's no such thing as full RDNA5 in latest Broken Silicon.

Well, he's right. AMD is the one who decides what to call RDNA2 and RDNA5, and AMD declared the PS5's iGPU is RDNA2. AMD never said the PS5 uses "RDNA1.5", and neither did they ever say "full RDNA2 is only GFX1020 and up". If AMD says the PS5 is RDNA2 and the PS5's ISA is GFX1013, then GFX1013 is RDNA2.

The "full RDNA2" thing was a marketing push by Microsoft (then massively adopted by Digital Foundry and some hardcore fans at Beyond3D, NeoGAF, etc.) to paint the Xbox Series' iGPUs as "more advanced" than the PS5's so it would give the idea that multiplatform games would look substantially better or run substantially faster on the Xbox and boost its sales.

Instead, none of that happened:

- DP4a capabilities of the Xbox consoles never got used for AI upscaling because there wasn't enough throughput for it and they ended up using regular temporal scaling like the PS5;
- Hardware Variable Rate Shading Tier 2 was so restrictive that every game/engine ended up using compute shader-based solutions with better granularity;
- Sampler Feedback Streaming only ever got used in 3dmark and some RTX demo because once again the practical results didn't match the theory so modern engines adopted software virtual texturing instead.

Games would perform neck and neck between the PS5 and the Series X, with the former even beating the latter in a couple scenarios where the additional fillrate and/or triangle output would compensate for the lower compute throughput and memory bandwidth.
And we're past the 5 year mark of those "longer" AAA development cycles, so if we haven't seen these adopted yet, then they're just not happening. Sony made the better decisions by locking down their specs earlier and focusing more on dev tools, as well as betting on narrow+fast instead of wide+slow for cheaper SoCs in the long run.




That said, if MLID saw internal documents describing the PS6's iGPUs as RDNA5, then it's RDNA5 regardless of ISA.

 

[MrMPFR]

Well, he's right. AMD is the one who decides what to call RDNA2 and RDNA5, and AMD declared the PS5's iGPU is RDNA2. AMD never said the PS5 uses "RDNA1.5", and neither did they ever say "full RDNA2 is only GFX1020 and up". If AMD says the PS5 is RDNA2 and the PS5's ISA is GFX1013, then GFX1013 is RDNA2.

It's still misleading regardless of how AMD spins it. The equivalent would be that the 7600 didn't have dual issue or work graphs HW and they still called it RDNA3.

No why would they. PS5 has V/F optimizations and RT units from RDNA2, the most important features.

Instead, none of that happened:

- DP4a capabilities of the Xbox consoles never got used for AI upscaling because there wasn't enough throughput for it and they ended up using regular temporal scaling like the PS5;
- Hardware Variable Rate Shading Tier 2 was so restrictive that every game/engine ended up using compute shader-based solutions with better granularity;
- Sampler Feedback Streaming only ever got used in 3dmark and some RTX demo because once again the practical results didn't match the theory so modern engines adopted software virtual texturing instead.

DP4a + XSX compute advantage makes FSR4 INT8 possible, if AMD ever decides to officially release it.

Upscaling made VRS completely irrelevant. It had some initial adoption in some NVIDIA sponsored titles and showed significant speedups.

NVIDIA RTX Texture Filtering in the HL2 RTX demo disagrees but maybe that just because the existing solution is crap?

Sampler feedback has two aspects. Another is TSS, which is very cool stuff + something AMD is investigating heavily ATM based on their patent filings.
A shame we never saw anyone bother, but why would they when PS5 doesn't support it.

Games would perform neck and neck between the PS5 and the Series X, with the former even beating the latter in a couple scenarios where the additional fillrate and/or triangle output would compensate for the lower compute throughput and memory bandwidth

Yeah XSX isn't very well balanced + has more bloated OS.

And we're past the 5 year mark of those "longer" AAA development cycles, so if we haven't seen these adopted yet, then they're just not happening. Sony made the better decisions by locking down their specs earlier and focusing more on dev tools, as well as betting on narrow+fast instead of wide+slow for cheaper SoCs in the long run.

No one bothered because main console didn't support any of it.

Also the PS5 Pro has the full DX12U equivalent feature set. Sony likely didn't opt to not use the extra functionality of RDNA2 dGPU and XSX, they prob just locked the feature set too early. If they thought this was irrelevant then why have they bothered to add the full feature set in the PS5 Pro.

That said, if MLID saw internal documents describing the PS6's iGPUs as RDNA5, then it's RDNA5 regardless of ISA.

Kepler already confirmed it's not PS5 vs XSX 2.0. For gaming it should have basically the full feature set, so any talk of missing features is splitting hairs.

 

[ToTTenTranz]

The equivalent would be that the 7600 didn't have dual issue or work graphs HW and they still called it RDNA3.

Don't know about work graphs hardware, but N33 does have dual issue:

 

[Kepler_L2]

Well, he's right. AMD is the one who decides what to call RDNA2 and RDNA5, and AMD declared the PS5's iGPU is RDNA2. AMD never said the PS5 uses "RDNA1.5", and neither did they ever say "full RDNA2 is only GFX1020 and up". If AMD says the PS5 is RDNA2 and the PS5's ISA is GFX1013, then GFX1013 is RDNA2.

The "full RDNA2" thing was a marketing push by Microsoft (then massively adopted by Digital Foundry and some hardcore fans at Beyond3D, NeoGAF, etc.) to paint the Xbox Series' iGPUs as "more advanced" than the PS5's so it would give the idea that multiplatform games would look substantially better or run substantially faster on the Xbox and boost its sales.

Instead, none of that happened:

- DP4a capabilities of the Xbox consoles never got used for AI upscaling because there wasn't enough throughput for it and they ended up using regular temporal scaling like the PS5;
- Hardware Variable Rate Shading Tier 2 was so restrictive that every game/engine ended up using compute shader-based solutions with better granularity;
- Sampler Feedback Streaming only ever got used in 3dmark and some RTX demo because once again the practical results didn't match the theory so modern engines adopted software virtual texturing instead.

Games would perform neck and neck between the PS5 and the Series X, with the former even beating the latter in a couple scenarios where the additional fillrate and/or triangle output would compensate for the lower compute throughput and memory bandwidth.
And we're past the 5 year mark of those "longer" AAA development cycles, so if we haven't seen these adopted yet, then they're just not happening. Sony made the better decisions by locking down their specs earlier and focusing more on dev tools, as well as betting on narrow+fast instead of wide+slow for cheaper SoCs in the long run.




That said, if MLID saw internal documents describing the PS6's iGPUs as RDNA5, then it's RDNA5 regardless of ISA.

What AMD says publicly and what they consider the PS5 to be internally are very different things. Upcoming CPUs with "Zen6LP" are another example where despite the name, it is considered "Zen5 with Zen6 ISA" internally.

 

[MrMPFR]

Don't know about work graphs hardware, but N33 does have dual issue:

View attachment 138842

They state 7000 series and newer in the Work graphs press releases + I know

I was just using it as an example to illustrate why calling the PS5 RDNA2 is misleading. Again I know PS5 has the most important stuff, but omitting mesh shaders, VRS, Sampler feedback (TSS and SFS) and DP4a isn't trivial, even if that functionality mostly hasn't been utilized.

 

[marees]

Upcoming CPUs with "Zen6LP" are another example where despite the name, it is considered "Zen5 with Zen6 ISA" internally.

ohhh

 

[ToTTenTranz]

What AMD says publicly and what they consider the PS5 to be internally are very different things.

The PS5 shows almost exactly the same performance-per-clock and relevant features as all other RDNA2 GPUs. It's a RT-enabled follow-up of RDNA1 that reaches significantly higher clocks, and the features it lacks from GFX1020+ ended up not being widely used. For all intents and purposes, what AMD said publicly seems more correct than whatever these internal discussions concluded.


I do remember a more valid dissonance from AMD, which was Vega M for Kaby Lake G. It was a Polaris chip with HBM2 and they even called the chip Polaris 22.

Again I know PS5 has the most important stuff, but omitting mesh shaders, VRS, Sampler feedback (TSS and SFS) and DP4a isn't trivial, even if that functionality mostly hasn't been utilized.

If it wasn't utilized, then it's trivial. That's pretty much what defines "trivial".

Upcoming CPUs with "Zen6LP" are another example where despite the name, it is considered "Zen5 with Zen6 ISA" internally.

Aren't the "LP" class of CPU cores an entirely new line developed mainly for lower leakage idle and OS tasks, though?

 

[Kepler_L2]

The PS5 shows almost exactly the same performance-per-clock

RDNA2 doesn't have any "IPC" gains

relevant features as all other RDNA2 GPUs.

No DP4a, no VRS, no Mesh Shaders, no Task Shaders, no SFS.

and the features it lacks from GFX1020+ ended up not being widely used.

Why would developers use a feature that the PS5 (with >75% market share) doesn't support?

Same thing happened with AMD on desktop PCs, they were first to many things like Tessellation, Async Compute, FP16 but those weren't used until NVIDIA supported it.

Aren't the "LP" class of CPU cores an entirely new line developed mainly for lower leakage idle and OS tasks, though?

They are derived from Zen5 though.

 

[MrMPFR]

If it wasn't utilized, then it's trivial. That's pretty much what defines "trivial".

Not trivial from a hardware perspective, I know adoption has been underwhelming to put it mildly for reasons explained by Kepler.

Mesh shaders are already close to primitive shaders on PS5, but lack of standardization (like where's task shader?) is problematic. It's extremely hard to program + chicken and egg situation on PC so not surprised we haven't seen adoption.
VRS is dead, SFS and TSS seems to have suffered same fate, but again not surprising. Last one is a big shame, from the NVIDIA blog on texture space shading:

TSS remembers which texels have been shaded and only shades those that have been newly requested. Texels shaded and recorded can be reused to service other shade requests in the same frame, in an adjacent scene, or in a subsequent frame. By controlling the shading rate and reusing previously shaded texels, a developer can manage frame rendering times, and stay within the fixed time budget of applications like VR and AR. Developers can use the same mechanisms to lower shading rate for phenomena that are known to be low frequency, like fog. The usefulness of remembering shading results extends to vertex and compute shaders, and general computations. The TSS infrastructure can be used to remember and reuse the results of any complex computation.

But seems like they're still trying to figure out how to make it work + like I said AMD is actively investigating this so it should be part of a major push with PS6. A shame a feature introduced with Turing and RDNA 2 likely won't see adoption until the time where these are completely irrelevant.

Also devs are not going to use DP4a for consoles when XSS can't run the stuff properly, Xbox is tiny portion of market + PS5 doesn't support it.
If it had DP4a then we would've seen a competent ML upscaler used PS5 on console at some point.

 

[MrMPFR]

Same thing happened with AMD on desktop PCs, they were first to many things like Tessellation, Async Compute, FP16 but those weren't used until NVIDIA supported it.

Wasn't there some BS with the God of War PC port or some other Sony exclusive on PC where they stripped out all the PS4 Pro optimizations because it negatively affected Paxwell cards?

no Task Shaders

Does RDNA 2 have this in HW or just some clever compiler rework?
Unfortunately IIRC it seems like no one can find a way to do it faster native vs just using a compute shader. I hope RDNA 5 does something to adress this.

Also from what I can read the RDNA 2 implementation of mesh shaders is quite bare bones, and not as good as the NVIDIA Turing one. Basically sounds like a low effort workaround to make primitive shaders work, although that could be wrong + info online is very limited.

RDNA 3 has two very important changes to how mesh shading is done, relating to vertex attribute export + adding wave-wide offset, as for the rest I'm not sure how it compares with NVIDIA. Not surprised they made a major push with Work graphs only 0.5 years after it launched. I'll also be interested to see what RDNA 5 does differently here (beyond the stuff already discussed) and how that impacts the PS6.

 

[ToTTenTranz]

RDNA2 doesn't have any "IPC" gains

Exactly. And for >95% of realistic use-cases, RDNA2 = RDNA1 + RT + higher clocks.

No DP4a, no VRS, no Mesh Shaders, no Task Shaders, no SFS.

Which are mostly irrelevant because they're not used.. Save for enabling DP4a XeSS on RDNA2 GPUs, neither PC nor Xbox hardware are making use of the rest.

Why would developers use a feature that the PS5 (with >75% market share) doesn't support?

1st party Xbox developers didn't use it either and Phil Spencer, Matt Booty, Sarah Bond, etc. could've dictated they had to use these to make a difference in their hardware.
SFS should've been a godsend for the Series S with its tiny 8GB max game allocation, yet it was completely absent from Starfield, Halo Infinite, Gears 5, CoD, Forza, etc. Neither at launch nor in an update.

The only "full RDNA2" feature we ever saw being used was VRS Tier 2 in Gears 5 after an update and that's it.


Microsoft is just as much responsible for not implementing the features in their software (for which they made a big marketing push) as Sony was for not implementing it in their hardware.

Also devs are not going to use DP4a for consoles when XSS can't run the stuff properly, Xbox is tiny portion of market + PS5 doesn't support it.
If it had DP4a then we would've seen a competent ML upscaler used PS5 on console at some point.

Microsoft is the biggest videogame publisher in the world and they couldn't mandate their own gamedevs to allocate the resources to develop a functional DP4a upscaler that runs on their ~30 million consoles?

 

[MrMPFR]

Microsoft is the biggest videogame publisher in the world and they couldn't mandate their own gamedevs to allocate the resources to develop a functional DP4a upscaler that runs on their ~30 million consoles?

Yeah well MS clearly doesn't care + XSS GPU is useless and weaker than a 6500XT, RX 470/1650 territory. Would've never worked when RX 6600 struggles with FSR4 INT8.

 

[marees]

i think the PS5 is a cutdown version of RDNA 2. not the real RDNA 2

For ex: FSR4/PSSR has a magical effect in RE Requiem

similar thing could have been attempted with int8 version if it had true RDNA 2

https://twitter.com/x/status/2026981801386819720

 

[MrMPFR]

Kepler has said we can expect ~9070XT raster for PS6.

What do y'all think in terms of performance expectations for ML, work graphs, neural shading, upscaler and denoiser, and path tracing?

 

[MrMPFR]

FSR4/PSSR has a magical effect in RE Requiem

Some speculation online that PSSR2 was actually used on RE Requiem. The wording DF used was very vague + some surprise video coming very soon.