• 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.

PS5 Pro Technical Seminar with Mark Cerny

DeepEnigma

Gold Member
Surely someone already answered this, but why is this not a problem on Windows?
Abstraction (derp) layers. In Windows, your extraction layers are higher up with less kernel access. PCs more often than not rely on brute forced methods for performance gains.

Consoles use a much lower extraction layer to get more out of a closed system ("punching above their weight"), thus the timings are more core dependent, etc.. "Coding to the metal." (meme)
 
Last edited:

cormack12

Gold Member
What did Mark just say?

9ebugx.gif
 
Last edited:

diffusionx

Gold Member
No, that's why I asked.



Surely someone already answered this, but why is this not a problem on Windows?



It's very much an overpriced product, but then again, that's really subjective...

You see a 2 year old GPU architecture and a 5 year old CPU as a "refresh" worth $800 and all I see is overpriced relics wrapped in a plastic box that shouldn't cost more than $600.

Why is there a need to compile a separate game required? There is no way this happens on PC the devs will go insane.
Implying pc devs aren’t already going insane, what do you shader stutter is coming from?
 
IjXMs6i.jpeg


So memory bandwidth was the bottleneck after all. Well done.
What else? It was so obvious.

The PS5 most likely just breaks even, given that Sony said they didn't have room to cut prices and Microsoft have admitted they were making a loss per unit. The extra BOM for the bigger APU and bigger SSD shouldn't be more than $150 and could be as low as $100, depending on what margins AMD charge. So we would expect Sony to be making at least $100 in profit per unit.
PS5 SSD is costing less than XSX SSD as it's fully custom and integrated to the MB. It's the RD costs that were higher but they likely reimbursed those costs already.
 

SweetTooth

Gold Member
What else? It was so obvious.


PS5 SSD is costing less than XSX SSD as it's fully custom and integrated to the MB. It's the RD costs that were higher but they likely reimbursed those costs already.

Funny thing is Cerny acknowledged it then spent a major chunk EXPLAINING the memory bottleneck and how they managed to cleverly solve it and what is the ideal situation for it and yet....

We get Slimy posts and other posters acting clever without watching the video 😅
 
Last edited:

Fafalada

Fafracer forever
Surely someone already answered this, but why is this not a problem on Windows?
It is a problem on Windows/PC.
Unlike consoles - you get runtime recompiling of shaders for every hardware permutation. This leads to both shader stutter and (often lengthy) shader precompilation steps at start - neither of which exist on consoles at all.
And utilization is also a problem - dual-issue has been universally treated as no-performance gain on PC. Basically - this one was the worst of both worlds scenario on PC.

Why is there a need to compile a separate game required?
See above (not the whole game - but the GPU code - yes). It's mostly users that are going insane over it though (and for good reasons).

Nope, NVIDIA architectures just have a single unified register file. It's actually AMD that has split register files (Vector+Scalar for RDNA, Vector+Scalar+Accumulation for CDNA).
So register remapping, but contending for same resources. Presumably PS5 Pro reuses one of the two register pools and didn't add another - which would make the whole use loosy comparable then (though I guess NVidia can get larger register file this way for the purpose).
 

sachos

Member
Forgive me for a diary style of post, but I absolutely envy Mark Cerny and his team's job.
I have this feeling not only with Mark's job but also AI labs. Imagine their feeling getting results of their latest massive secret training runs, must be amazing.
I think he said that it would be labeled differently and not be PSSR. Almost like PSSR is this version and next will be a different name vs PSSR 2.0
I think by that he meant that "Playstation Spectral" is their branding for AI models and the other two letters would be the technique/model, so Frame Gen would be PSFG if they follow that logic. But i did not get the feeling that they would stop naming "PSSR" PSSR if they keep on improving it.
 

yogaflame

Member
In the early stages of PSSR ML for Ps5 pro, I hope all developers will take the the option of having an option of 40 -50 fps with those with RT/4k ( especially with UE5 ) just quick ps5 pro update for now and by next year once PSSR ML learned more and matured, they can gradually implement 60fps with 4k/RT.
 
Last edited:

Bojji

Member
like what features? Can you see it?
or you just follow the nonsense on social media?

thats like sayinh "this game runs at 1200 p only, it should run at 1440p, yea i can see it, i can see it!

VRS, SFS, Mesh Shaders and DP4a support (XeSS could run on this).

Is this important? Sony clearly thinks so because Pro has all this - and they risked breaking that "one .exe" file compatibility for PS5 and Pro Cerny was talking about.

PS5 has RDNA1 raster core and RDNA2 RT - Just like Pro is RDNA 2.X for raster and RDNA4 for RT.
 

Three

Gold Member
VRS, SFS, Mesh Shaders and DP4a support (XeSS could run on this).

Is this important? Sony clearly thinks so because Pro has all this - and they risked breaking that "one .exe" file compatibility for PS5 and Pro Cerny was talking about.

PS5 has RDNA1 raster core and RDNA2 RT - Just like Pro is RDNA 2.X for raster and RDNA4 for RT.
How do you go from saying TOPS performance isn't important to now saying these things were important despite years of proof that they weren't?
 
Last edited:

SKYF@ll

Member
Cerny talk didn't change the fact that PS5 is missing RDNA2 features.
RDNA2 features. *XBOX only
HW VRS: Only slight performance improvement with significant degradation in IQ.
Sampler Feedback Streaming: Tech Demo Only
Mesh Shader: As for HW, Primitive Shader still exists( PS5, XSS/X , RDNA2-3 ), and the usage of Mesh Shader can be realized with Primitive Shader.

PS5 custom features.
ID Buffer (for CBR)
Foveated Rendering (for VR)
Ray Tracing : Early subtree culling , PS5 can allow spilling state to either LDS or scratch or just kept in registers.
*RDNA2 Late culling *On Xbox the state goes into scratch memory.
Boost Clock : Fine-grained clock optimization for each functional block in the GPU chip (clock gating optimization)
Smart Shift : Dynamically change power allocation to APU/dGPU depending on workload
 

Three

Gold Member
Extraction layers. In Windows, your extraction layers are higher up with less kernel access. PCs more often than not rely on brute forced methods for performance gains.

Consoles use a much lower extraction layer to get more out of a closed system ("punching above their weight"), thus the timings are more core dependent, etc.. "Coding to the metal." (meme)
Abstraction layers. Unless we're talking about something else actually called extraction layers?
 

Crayon

Member
You know what part of the reason that's fun to listen to guy is? They asked him what spectral is supposed to mean, and instead of going into a bunch of wishy-washy s*** about spectrums and bringing things together and ideals and blah blah blah, he just says we made it up it doesn't mean anything.
 

LordOfChaos

Member
You know what part of the reason that's fun to listen to guy is? They asked him what spectral is supposed to mean, and instead of going into a bunch of wishy-washy s*** about spectrums and bringing things together and ideals and blah blah blah, he just says we made it up it doesn't mean anything.
Yup, total class act, never overhyped, just explained like an engineer
 

SKYF@ll

Member
VRS, SFS, Mesh Shaders and DP4a support (XeSS could run on this).

Is this important? Sony clearly thinks so because Pro has all this - and they risked breaking that "one .exe" file compatibility for PS5 and Pro Cerny was talking about.

PS5 has RDNA1 raster core and RDNA2 RT - Just like Pro is RDNA 2.X for raster and RDNA4 for RT.
PS5: RDNA 2 raster + RDNA2 RT + RDNA1 Front-End +RDNA1 Render-Back-Ends
XSX: RDNA 2 raster + RDNA2 RT + RDNA1 Front-End +RDNA2 Render-Back-Ends
PS5 Pro: RDNA 2 raster (RDNA 3 Geometry Engine) + RDNA4 RT + 300TOPS(8bit) , 66.7TFLOPS(16bit) for AI+ RDNA2 Render-Back-Ends
GiTo5XH.jpg
 

Bojji

Member
How do you go from saying TOPS performance isn't important to now saying these things were important despite years of proof that they weren't?

I never said they were, I said that Sony thinks they are because they didn't gut them from pro GPU.

VRS is garbage, mesh shaders and primitive shaders (supported by PS5) are more or less the same.

Returning to why this discussing even started: PS5 lacks rDNA 2 features. Xbox fanboys called it 1.5 to mock it but in the end it's close to reality. Did it stop console from being more or less on par with XSX? So far, it did not.
 
Last edited:

Three

Gold Member
I never said they were, I said that Sony thinks they are because they didn't gut them from pro GPU.
Yet the TOPs/ML performance they included isn't? Even though they relate to some of the things you list?
VRS is garbage, mesh shaders and primitive shaders (supported by PS5) are more or less the same.

Returning to why this discussing even started: PS5 lacks rDNA 2 features. Xbox fanboys called it 1.5 to mock it but in the end it's close to reality. Did it stop console from being more or less on par with XSX? So far, it did not.
If this is the case why did you see the need to reply to this very simple post

"Also I like how PS5 was RDNA1.x talk disappeared"

By emphasising "PS5 is missing features".
Get told it meant f' all really.
Then come back with "it's important otherwise they wouldn't include it" when you were arguing these included things weren't important. I don't get it.
 

Bojji

Member
Yet the TOPs/ML performance they included isn't? Even though they relate to some of the things you list?

If this is the case why did you see the need to reply to this very simple post

"Also I like how PS5 was RDNA1.x talk disappeared"

By emphasising "PS5 is missing features".
Get told it meant f' all really.
Then come back with "it's important otherwise they wouldn't include it" when you were arguing these included things weren't important. I don't get it.

I think they aren't that much important. At the same time they are part of RDNA2 features, PS5 is missing quite some stuff from RDNA2.

Things that are different between RDNA1-RDNA2: VRS, SFS, RT, DP4A support, infinity cache.

PS5 only has RT, while XSX and Pro have all of it minus infinity cache. So why are people angry when someone says that PS5 isn't RDNA2?
 

Three

Gold Member
I think they aren't that much important. At the same time they are part of RDNA2 features, PS5 is missing quite some stuff from RDNA2.

Things that are different between RDNA1-RDNA2: VRS, SFS, RT, DP4A support, infinity cache.

PS5 only has RT, while XSX and Pro have all of it minus infinity cache. So why are people angry when someone says that PS5 isn't RDNA2?
Nobody is angry. I think you're just often missing the context of the conversation. He mentioned the talks of RDNA1.x disappearing because twilo was making the same mistake again of reintroducing "not full RDNA3". Then when you swooped in to reply and tell everyone base PS5 is missing features of "full RDNA2" he clarified by saying look at the actual results of this custom RDNA in PS5 base, how those things were meaningless/unimportant overall.
You then replied about how it's actually important, otherwise they wouldn't have include it now? While earlier you were saying the included boost in AI/TOPS performance on a PS5 Pro was not important because devs don't use those things like mesh shaders or dp4a?
To me it appears like you're just belittling engineering decisions in a somewhat contradictory manor.
 

Fafalada

Fafracer forever
Things that are different between RDNA1-RDNA2: VRS, SFS, RT, DP4A support, infinity cache.
Not to be the grammar nazi here - but there's no such feature as SFS, in any architecture(or DX12 spec).
Sampler Feedback (no S) is a GPU feature, and I'm not sure anyone has ever looked at whether PS5 supports it or not.

And Microsoft talked at length about how to implement Sampler Feedback Streaming (the algorithm), using their custom extensions(not part of RDNA2) that they added to SF.

</pedantic rant>
 

PaintTinJr

Member
Nobody is angry. I think you're just often missing the context of the conversation. He mentioned the talks of RDNA1.x disappearing because twilo was making the same mistake again of reintroducing "not full RDNA3". Then when you swooped in to reply and tell everyone base PS5 is missing features of "full RDNA2" he clarified by saying look at the actual results of this custom RDNA in PS5 base, how those things were meaningless/unimportant overall.
You then replied about how it's actually important, otherwise they wouldn't have include it now? While earlier you were saying the included boost in AI/TOPS performance on a PS5 Pro was not important because devs don't use those things like mesh shaders or dp4a?
To me it appears like you're just belittling engineering decisions in a somewhat contradictory manor.
The double standard seems to be driven by: 'Not the way Nvidia is doing it, so it is wrong!' and that's felt like the angle since the: 'Not full RDNA2' projection at the start of the gen yielded negative benefits against the far, far superior PS5 APU design.

IMO this technical talk just highlights a recurring approach by Sony and AMD from their collaboration, the end goal always seems to be maximising silicon use for maximum utilisation for all rendering paradigms at the same time, as illustrated by the Pro having 300TOPs for ML and unbelievable stacked memory bandwidth for the CNNs, but even when it is typically utilised for PSSR's 1ms a frame leading to 3%-12% per second usage the remaining capacity is filled with rasterization, GPU Compute or RT workloads, embracing the ethos of the RISC-CISC hybrid, and long term this design philosophy of maximizing die space at optimal power will eventually result in a final destination that even Nvidia's market leading position by outspending on bigger dies with faster memory and more transistors will struggle to live with - unless they change their philosophy IMO.
 
Last edited:

Zathalus

Member
Probably find out soon. Even Nvidia can't think they can keep brute forcing it though, so expect them to instead find ways to use the tensor cores as it seems a lot of potential is left sitting on the table, albeit used intensively for brief bursts.
Nvidia is not just brute forcing things though. In terms of performance per watt and efficiency of die space they still lead both Intel and AMD. It just so happens they take that and throw enormous amounts of power and a massive die to get the absolutely peak performance for the 90 series cards, but the majority of their offerings are much smaller chips. AD106 is a mere 188mm2 for example.
 

PaintTinJr

Member
Nvidia is not just brute forcing things though. In terms of performance per watt and efficiency of die space they still lead both Intel and AMD. It just so happens they take that and throw enormous amounts of power and a massive die to get the absolutely peak performance for the 90 series cards, but the majority of their offerings are much smaller chips. AD106 is a mere 188mm2 for example.
They don't lead in raster in 'relative' terms going by twitch shooters like CoD's performance on AMD cards AFAIK. They have brute forced DLSS by having a subsystem that sits idle between the large window of upscaling, and they brute force RT also, by having a dedicated subsystem that sits idle when RT isn't used, unless I'm misunderstanding their designs.

edit:
And I was under the impression they always use more cache at faster rates and faster VRAM and/or higher interface wiring counts, and lower relative lithography nodes - they can do cheaper as market leader - which is again just outspending the competition unless I'm missing something
 
Last edited:
RDNA2 features. *XBOX only
HW VRS: Only slight performance improvement with significant degradation in IQ.
Sampler Feedback Streaming: Tech Demo Only
Mesh Shader: As for HW, Primitive Shader still exists( PS5, XSS/X , RDNA2-3 ), and the usage of Mesh Shader can be realized with Primitive Shader.

PS5 custom features.
ID Buffer (for CBR)
Foveated Rendering (for VR)
Ray Tracing : Early subtree culling , PS5 can allow spilling state to either LDS or scratch or just kept in registers.
*RDNA2 Late culling *On Xbox the state goes into scratch memory.
Boost Clock : Fine-grained clock optimization for each functional block in the GPU chip (clock gating optimization)
Smart Shift : Dynamically change power allocation to APU/dGPU depending on workload
PS5 has a few others custom features. Also ID Buffer can be used in some TAA implementations to remove ghosting.
PS5: RDNA 2 raster + RDNA2 RT + RDNA1 Front-End +RDNA1 Render-Back-Ends
XSX: RDNA 2 raster + RDNA2 RT + RDNA1 Front-End +RDNA2 Render-Back-Ends
PS5 Pro: RDNA 2 raster (RDNA 3 Geometry Engine) + RDNA4 RT + 300TOPS(8bit) , 66.7TFLOPS(16bit) for AI+ RDNA2 Render-Back-Ends
GiTo5XH.jpg
Great summary (and your previous posts) but how do we know PS5 Pro has RDNA2 Render-backs-ends?
PS5 Pro RDNA2 Render-Back-Ends
 
Last edited:

Zathalus

Member
They don't lead in raster in 'relative' terms going by twitch shooters like CoD's performance on AMD cards AFAIK. They have brute forced DLSS by having a subsystem that sits idle between the large window of upscaling, and they brute force RT also, by having a dedicated subsystem that sits idle when RT isn't used, unless I'm misunderstanding their designs.
If you compare die space vs die space then Nvidia usually comes out ahead. For example the 7800 XT is 346 mm² vs the 294 mm² of the 4070 Ti. Power usage is similar between the two but the 4070 Ti is usually faster and is a smaller chip. In the latest COD the 4070ti and 7800XT basically tie, but that 4070Ti chip is still smaller (with far less memory bandwidth as well). Another pure raster game that launched recently, Space Marine 2, has the 4070 Ti leasing by 23%. AMD of course wins when it comes to price, as Nvidia are charging a premium they know they can get away with.

Tensor and RT cores sit idle if they are not being used, but the die space they take up is relatively small vs the entire GPU so the trade-off seems worth it (for now).
 
Last edited:

SKYF@ll

Member
Great summary (and your previous posts) but how do we know PS5 Pro has RDNA2 Render-backs-ends?

PS5 Pro Runs Naraka: Bladepoint Twice as Fast, Says Developer; PSSR Has Similar Quality to DLSS​

"Naraka: Bladepoint can achieve 4K 60FPS on the PS5 Pro without using upscaling technologies, which is not possible on the PS5."
"The most impressive features are PSSR and VRS. In particular, VRS, when applied to specific rendering processes in Naraka: Bladepoint, can drastically reduce rendering time. "

These are the words of the developers of Naraka.
I have never played this game, but I think it is a free game.
 

PS5 Pro Runs Naraka: Bladepoint Twice as Fast, Says Developer; PSSR Has Similar Quality to DLSS​

"Naraka: Bladepoint can achieve 4K 60FPS on the PS5 Pro without using upscaling technologies, which is not possible on the PS5."
"The most impressive features are PSSR and VRS. In particular, VRS, when applied to specific rendering processes in Naraka: Bladepoint, can drastically reduce rendering time. "

These are the words of the developers of Naraka.
I have never played this game, but I think it is a free game.
But he is talking about software features, he is not saying ML hardware and RDNA2 VRS.

PSSR on its own is a software and could run on anything and VRS can be done 100% without specific hardware (and can give better results that way as shown in a famous COD Slide). His developers are maybe using VRS in that game, but it doesn't mean anything about the hardware behind. VRS doesn't mean RDNA2 VRS
 

Bojji

Member
Nobody is angry. I think you're just often missing the context of the conversation. He mentioned the talks of RDNA1.x disappearing because twilo was making the same mistake again of reintroducing "not full RDNA3". Then when you swooped in to reply and tell everyone base PS5 is missing features of "full RDNA2" he clarified by saying look at the actual results of this custom RDNA in PS5 base, how those things were meaningless/unimportant overall.
You then replied about how it's actually important, otherwise they wouldn't have include it now? While earlier you were saying the included boost in AI/TOPS performance on a PS5 Pro was not important because devs don't use those things like mesh shaders or dp4a?
To me it appears like you're just belittling engineering decisions in a somewhat contradictory manor.

They don't appear very important because small fraction of games use them. Why? Maybe because most popular console (target platform) doesn't support it? When it comes to what I think about those features:

VRS - dogshit
Mesh Shaders - Very important (in the long run) but PS5 supports Primitive Shaders so it doesn't change much for this console
SFS - who knows
No DP4a support - this killed potential for ML uspacling being used on regular PS5 (while you can run XESS on RDNA2 AMD GPUs on PC)



Tensor Cores/Custom RDNA hardware in Pro - ML hardware is something that will be important in the future for game rendering but IMO not for next few years. Right now they pretty much only use it for SR (DLSS, PSSR, XeSS).

PS5 Pro Runs Naraka: Bladepoint Twice as Fast, Says Developer; PSSR Has Similar Quality to DLSS​

"Naraka: Bladepoint can achieve 4K 60FPS on the PS5 Pro without using upscaling technologies, which is not possible on the PS5."
"The most impressive features are PSSR and VRS. In particular, VRS, when applied to specific rendering processes in Naraka: Bladepoint, can drastically reduce rendering time. "

These are the words of the developers of Naraka.
I have never played this game, but I think it is a free game.

Dear developers, take that VRS and...

giphy.webp
 

PaintTinJr

Member
If you compare die space vs die space then Nvidia usually comes out ahead. For example the 7800 XT is 346 mm² vs the 294 mm² of the 4070 Ti. Power usage is similar between the two but the 4070 Ti is usually faster and is a smaller chip. In the latest COD the 4070ti and 7800XT basically tie, but that 4070Ti chip is still smaller (with far less memory bandwidth as well). Another pure raster game that launched recently, Space Marine 2, has the 4070 Ti leasing by 23%. AMD of course wins when it comes to price, as Nvidia are charging a premium they know they can get away with.

Tensor and RT cores sit idle if they are not being used, but the die space they take up is relatively small vs the entire GPU so the trade-off seems worth it (for now).
But you can't you would have to compare spend on the die, so it is 'relative' in a true sense and not being able to offer interface connections, cache and vram and lithography levels that the competition can't. Ultimately Nvidia produce nothing that will be more power efficient and performant as the Pro, at the same size, and that's despite half the die space in a Pro being used by a CPU.
 
Last edited:

PaintTinJr

Member
So you picked the one game where AMD has a significant advantage to make a broad generalization. Did I get this right?
I used the one that has been refined and evolved over decades, where bad port accusations, or not pushing visuals can't be used, and being the biggest game in town would run better on Nvidia at pure raster if it possibly could given Nvidia own the dedicated card market and running best on Nvidia doesn't hurt the game's marketing.

edit:
And we know for certain that performance is the priority in the development, so it isn't a heavy deferred renderer hiding memory latency issues for one or the other but trying to be as fast a forward renderer/+ as can be.
 
Last edited:

Clear

CliffyB's Cock Holster
The thing that should be taken away from this discussion, and applied generally to all related subjects, is the paramount importance of system I/O.

It doesn't matter how fast the internals are on any component if there's bottleneck to data coming in or out of it,.
 
Top Bottom