Support NeoGAF

[Riky]

He still doesn't understand that XBSX/S is the only console that supports DirectX 12 Ultimate features. Features the PS5 doesn't support because it uses it's own API features.

It's supported in Vulcan.

 

[azertydu91]

What?

Yep they did and even reiterated after the ps5 presentation where it was explicitely stated that it had raytracing....But that's battaglia there, you can watch their reaction to the ps5 reveal it is inside.

 

[Max Payne’s Baretta]

You either have technical chops or ya don’t. Clear to see who doesn’t due to this thread.

 

[MonarchJT]

I literally posted exactly what Cerny said, he is the lead architect of the PS5.

He said Primitive Shaders in the PS5 is a brand new capability that vary the details of objects and effects. Something much different from what AMD Primitive Shaders does.

The Geometry Engine on the other does exactly the same thing as Mesh Shaders.

He also did a patent describing how VRS is done using Primitive Shaders.

The guy literally made an insane I/O unit in the GPU that removes bottlenecks and repurposed a compute unit into the Tempest Engine.
He can also re-enigineer the Geometry Engine in the PS5 to do what he needs it to do.

But it doesn't matter because you Xbots are bewitched behind Microsoft.

this is becoming misterxmedia level
someone call the ambulance

 

[thicc_girls_are_teh_best]

Every technical thread now turns into a show, the hell? Sony Ponies taking every dig at Xbox, Xbutts taking every dig at PlayStation, meanwhile Nintendrones are probably disguised as both playing both sides and you can't see it because of the makeup running down the face.

Do people really need to keep using these threads to list bullet points justifying teh powwah of their favorite piece of plastic?

 

[Tripolygon]

How about you all stop arguing about technologies that have not shown up in any significant way in these consoles. You are arguing without actually having seen how it applies to consoles. There is only one console game so far that has shown the use of deep learning inference and that is spiderman for real-time muscle deformation. VRS is a performance-saving technology that all GPUs can use with varying implementation which trades image quality for performance. How are Meshshaders/Primitive Shaders being currently used in these consoles?

It is time to show examples of how it significantly benefits games.

 

[Loxus]

this is becoming misterxmedia level
someone call the ambulance

How the fuck is using information from Mark Cerny comparable to that. It isn't my fault you can't understand a basic concept.

Primitive Shaders does VRS in the early stages of the rendering pipeline. It's described in one of Mark Cerny's patents.

I also provided information from Road to PS5 that Primitive Shaders in the PS5 can vary the level of details and effects, while AMD Primitive Shaders can not.

You don't provide shit to prove me wrong but fanboyism.

 

[MonarchJT]

How the fuck is using information from Mark Cerny comparable to that. It isn't my fault you can't understand a basic concept.

Primitive Shaders does VRS in the early stages of the rendering pipeline. It's described in one of Mark Cerny's patents.

I also provided information from Road to PS5 that Primitive Shaders in the PS5 can vary the level of details and effects, while AMD Primitive Shaders can not.

You don't provide shit to prove me wrong but fanboyism.

wtf you talking about ....mesh shaders or the old primitive shader do culling ....at the initial stage
vrs must be applied at the end of the pipeline ..and it "shade" with more or less details the geometry passed by the initial stage

you really have no idea of what you talking about and with this I'm out

 

[winjer]

wtf you talking about ....mesh shaders or the old primitive shader do culling ....at the initial stage
vrs must be applied at the end of the pipeline ..and it "shade" with more or less details the geometry passed by the initial stage

you really have no idea of what you talking about and with this I'm out

The patent he is talking about, starts by describing the process of rasterization, which means it talks about Primitives (points, lines and triangles)
But because he doesn't understand anything, he thinks it's talking about "Primitive Shaders" (geometry pipeline).
He is just mixing up two different terms.

 

[MonarchJT]

The patent he is talking about, starts by describing the process of rasterization, which means it talks about Primitives (points, lines and triangles)
But because he doesn't understand anything, he thinks it's talking about "Primitive Shaders" (geometry pipeline).
He is just mixing up two different terms.

exactly

 

[IllegalLemon]

The thing that baffles me, is why people in here are arguing about graphics technology that only graphics engineers know the ins and outs of?
The question that remains in my mind for everyone in this thread is: Have you people used and coded for the devkits of both consoles, to be able to tell which is more efficient, and at which graphics processing workload? Do you know of what goes in the graphics processing hardware & engines of both SOC's that makes them custom?

If you don't, then why are you still arguing about things only a graphics engineer/programmer would know? I mean, I don't know about the different tricks and customisations that consoles have under the hood, and I wouldn't go out of my way to tell which is better. This makes me think if this thread was created for the sole purpose of inciting the console wars and has so far lead to nothing but dead ends, turning into a absolute dumpster fire.

So the question remains. Have you people used these technologies to be able to deduce which one works best, and at which scenario? Or are you basing your claims purely on marketing speak, research papers and vague power point slides? If you haven't used them, or seen them implemented in games in any way, then why are you still arguing about it and over things both you and I have practically no deep understanding of whatsoever. What is this thread(whose purpose was taken completely out of context) and all this arguing trying to achieve? Because I noticed that it's getting nowhere and it's becoming more of an annoyance than anything meaningful.

 

[ToTTenTranz]

This makes me think if this thread was created for the sole purpose of inciting the console wars and has so far lead to nothing but dead ends, turning into a absolute dumpster fire.

Bingo.


If that wasn't obvious from the start, here's the OP in the first post:

Luckily for (Series gamers) it looks like this technology will be supported by the Series consoles due to hardware support for Int-4 dot operations.

Here's what Intel themselves had to say about the open source XeSS he's talking about:

You've seen these machine learning style applications run on GPUs with none, right? There's no reason it has to have a particular hardware.

Finally, if you check OP's posts throughout the rest of the thread, they consist mostly of one-liners insulting other users or claiming how everyone calling him out is so ignorant.

It's a flamebait subject with a flamebait thread title and flamebait posts.
This wasn't going to end up as anything other than a shitshow.

 

[Loxus]

The patent he is talking about, starts by describing the process of rasterization, which means it talks about Primitives (points, lines and triangles)
But because he doesn't understand anything, he thinks it's talking about "Primitive Shaders" (geometry pipeline).
He is just mixing up two different terms.

This is exactly how Mark Cerny described Primitive Shaders in the PS5.

More complex usage involves something called primitive shaders which allow the game to synthesize geometry on-the-fly as it's being rendered.
It's a brand new capability.
Using primitive shaders on PlayStation 5 will allow for a broad variety of techniques including smoothly varying level of detail, addition of procedural detail to close up objects and improvements to particle effects and other visual special effects.

This is how the patent works.

As simple as I can put it, basically what the patents are saying. Geometry is grouped in sections similarly to how nanite group triangles into clusters. This whole process is what Primitive Shaders is doing, synthesize geometry on the fly.
Synthesize - to combine (a number of things) into a coherent whole.
I wouldn't be surprised if Primitive Shaders are also responsible for running Nanite.

All of this is done in the Geometry Stage. Then these geometry sections are sent through the rasterizer and saved in cache and memory in the form of metadata. This metadata contains information that determines how many pixels to generate in each section.
Metadata - a set of data that describes and gives information about other data.

Pixel Shaders then access this metadata to generate the final pixels in each of the sections to have different pixel resolutions.

The patent also talked about overlapping sections. This is where Primitive Shaders come in again. It varies the levels of details in the sections. The results in varying levels of pixel are as seen below.

This technique works like Nanite but instead of triangles it's pixels.

The process is different but you get the same results as VRS.

GPUs have a component called pixel shaders, which are each assigned a pixel to dictate its visual characteristics. More detail calls for more graphics power and vice versa.

VRS changes the number of pixels one pixel shader operation can impact. As a result, one pixel shader operation can be applied to a whole area of pixels -- a 16 x 16 pixel area, to be exact.

So like I said, VRS starts in the early stages of the rendering pipeline. This also coincides with what Matt Hargett was saying.

I don't expect you to comprehend any of this. You don't even understand the Geometry Engine basic functions is the same as Mesh Shaders.

PlayStation 5 has a new unit called the Geometry Engine which brings handling of triangles and other primitives under full programmatic control. Simple usage could be performance optimizations such as removing back faced or off-screen vertices and triangles.

 

[MonarchJT]

This is exactly how Mark Cerny described Primitive Shaders in the PS5.

More complex usage involves something called primitive shaders which allow the game to synthesize geometry on-the-fly as it's being rendered.
It's a brand new capability.
Using primitive shaders on PlayStation 5 will allow for a broad variety of techniques including smoothly varying level of detail, addition of procedural detail to close up objects and improvements to particle effects and other visual special effects.

This is how the patent works.

As simple as I can put it, basically what the patents are saying. Geometry is grouped in sections similarly to how nanite group triangles into clusters. This whole process is what Primitive Shaders is doing, synthesize geometry on the fly.
Synthesize - to combine (a number of things) into a coherent whole.
I wouldn't be surprised if Primitive Shaders are also responsible for running Nanite.

All of this is done in the Geometry Stage. Then these geometry sections are sent through the rasterizer and saved in cache and memory in the form of metadata. This metadata contains information that determines how many pixels to generate in each section.
Metadata - a set of data that describes and gives information about other data.

Pixel Shaders then access this metadata to generate the final pixels in each of the sections to have different pixel resolutions.

The patent also talked about overlapping sections. This is where Primitive Shaders come in again. It varies the levels of details in the sections. The results in varying levels of pixel are as seen below.

This technique works like Nanite but instead of triangles it's pixels.

The process is different but you get the same results as VRS.

GPUs have a component called pixel shaders, which are each assigned a pixel to dictate its visual characteristics. More detail calls for more graphics power and vice versa.

VRS changes the number of pixels one pixel shader operation can impact. As a result, one pixel shader operation can be applied to a whole area of pixels -- a 16 x 16 pixel area, to be exact.

So like I said, VRS starts in the early stages of the rendering pipeline. This also coincides with what Matt Hargett was saying.

I don't expect you to comprehend any of this. You don't even understand the Geometry Engine basic functions is the same as Mesh Shaders.

PlayStation 5 has a new unit called the Geometry Engine which brings handling of triangles and other primitives under full programmatic control. Simple usage could be performance optimizations such as removing back faced or off-screen vertices and triangles.

the tweet from Matt should enlighten you on the pipeline....GE do the culling and after come vrs in act ......in fact he writes that without the GE you would find yourself managing geometries that you could have eliminated in a first phase (precisely culling)
I'm sure than the revisited primitive shader of the PS5 is more flexible but i'm not at all sure it uses meshlet....or Cerny wouldn't have talked about "primitive" shaders at all

 

[winjer]

omg, this is almost to silly to be true.

 

[Barlow]

Reading this again feels like we're stuck in 2019/2020

 

[Epic Sax CEO]

This is exactly how Mark Cerny described Primitive Shaders in the PS5.

More complex usage involves something called primitive shaders which allow the game to synthesize geometry on-the-fly as it's being rendered.
It's a brand new capability.
Using primitive shaders on PlayStation 5 will allow for a broad variety of techniques including smoothly varying level of detail, addition of procedural detail to close up objects and improvements to particle effects and other visual special effects.

This is how the patent works.

As simple as I can put it, basically what the patents are saying. Geometry is grouped in sections similarly to how nanite group triangles into clusters. This whole process is what Primitive Shaders is doing, synthesize geometry on the fly.
Synthesize - to combine (a number of things) into a coherent whole.
I wouldn't be surprised if Primitive Shaders are also responsible for running Nanite.

All of this is done in the Geometry Stage. Then these geometry sections are sent through the rasterizer and saved in cache and memory in the form of metadata. This metadata contains information that determines how many pixels to generate in each section.
Metadata - a set of data that describes and gives information about other data.

Pixel Shaders then access this metadata to generate the final pixels in each of the sections to have different pixel resolutions.

The patent also talked about overlapping sections. This is where Primitive Shaders come in again. It varies the levels of details in the sections. The results in varying levels of pixel are as seen below.

This technique works like Nanite but instead of triangles it's pixels.

The process is different but you get the same results as VRS.

GPUs have a component called pixel shaders, which are each assigned a pixel to dictate its visual characteristics. More detail calls for more graphics power and vice versa.

VRS changes the number of pixels one pixel shader operation can impact. As a result, one pixel shader operation can be applied to a whole area of pixels -- a 16 x 16 pixel area, to be exact.

So like I said, VRS starts in the early stages of the rendering pipeline. This also coincides with what Matt Hargett was saying.

I don't expect you to comprehend any of this. You don't even understand the Geometry Engine basic functions is the same as Mesh Shaders.

PlayStation 5 has a new unit called the Geometry Engine which brings handling of triangles and other primitives under full programmatic control. Simple usage could be performance optimizations such as removing back faced or off-screen vertices and triangles.

I get it, but maybe you're missing a point?
PS5 Primitive Shaders IS NOT VRS, NO MATTER IF IT'LL ACHIEVE THE SAME RESULT!
SeX have the advantage that is can use VRS (even if badly) to save a bit of performance on current/traditionally made games, something that should be useful for the PS5 with its lower compute (leaving more to RT?). To take advantage of this potential games will need new engines fully made with Mesh Shaders, and when this happens... the SeX will also benefit!
Also, it's good and all that Cerny got the RDNA1 hardware and improved, but you forget that AMD also did the same for RDNA2? There's a Mesh Shaders benchmark for PC and if I'm remembering right while it works on RDNA1 and the performance improvement is huge, it's noticibly better on RNDA2, meaning that the SeX should easily outperform the PS5 and then PS5's Pixel and Discard advantage will not make the same difference anymore.

Naturally during the course of their lives we'll see better looking and better performing games (in part due to developers creativity finding solutions to problems and limitations), but one of these consoles have the potential to come ahead and it's not the PS5 (unfortunately).

 

[HoofHearted]

 

[oldergamer]

Why are people still entertaining loxus, aka misterxmedia?

 

[skneogaf]

Seems like the xbox series consoles have all the right parts to be the clear leader but needs years to get everything working at the same time.

 

[Loxus]

I get it, but maybe you're missing a point?
PS5 Primitive Shaders IS NOT VRS, NO MATTER IF IT'LL ACHIEVE THE SAME RESULT!
SeX have the advantage that is can use VRS (even if badly) to save a bit of performance on current/traditionally made games, something that should be useful for the PS5 with its lower compute (leaving more to RT?). To take advantage of this potential games will need new engines fully made with Mesh Shaders, and when this happens... the SeX will also benefit!
Also, it's good and all that Cerny got the RDNA1 hardware and improved, but you forget that AMD also did the same for RDNA2? There's a Mesh Shaders benchmark for PC and if I'm remembering right while it works on RDNA1 and the performance improvement is huge, it's noticibly better on RNDA2, meaning that the SeX should easily outperform the PS5 and then PS5's Pixel and Discard advantage will not make the same difference anymore.

Naturally during the course of their lives we'll see better looking and better performing games (in part due to developers creativity finding solutions to problems and limitations), but one of these consoles have the potential to come ahead and it's not the PS5 (unfortunately).

So your saying Cerny made that patent for nothing?

Did you even read how the patent does it?

 

[Loxus]

Why are people still entertaining loxus, aka misterxmedia?

Please explain to me how is Cerny like misterxmedia?

The patent made by Cerny explain how it varies pixels on screen.

Did you read the patent or are you in denial?

 

[Locuza]

| Answer Part 1: I always got an error message, without specifying the issue, when I tried to post everything together, the cut was made till the error point |

Why isnt the nVidia 1070GTX listed?

For space reasons I did not list every individual product based on a specific chip, but gave just one SKU example.
The GP104 chip is used by the 1070, 1070 Ti, 1080 GTX, but also some 1060 GPU models (+ mobile and professional GPUs).

I'm honestly not that sure about "democratization", as Intel's model still needs proper dedicated hardware to work at full power (i.e. providing useful performance or quality increase). It's the same with Nvidia - they could make DLSS open-source tomorrow and it won't change a thing, the performance hit from running DLSS without matrix-multiplication processors like Tensor cores will make the tech all but worthless unless you already have ridiculously strong hardware. Or some other type of dedicated ML accelerator, which is probably not going to be as good a solution either.

That's actually an open question.
According to Intel:

Karthik: Yes, we require inference capabilities but matrix acceleration is not the only form of inference capability that is available on GPUs. If you go all the way back to- I think Skylake- we had dot product acceleration, which is DP 4.4 – there’s various names for it. Nvidia has had this I think, since Turing and AMD has this now on RDNA2. So even without Matrix acceleration you can go quite far. It might not be as fast as matrix acceleration, but certainly meets the objective. And as I said, the objective is to maintain the fidelity of your render and achieve smooth frame rates. So, when it comes to older models, on older internal GPUs, we've had dot product acceleration (DP4a) for a while now.
Microsoft has enabled this through Shader Model 6.4 and above and on all these platforms XeSS will work.

https://wccftech.com/intel-xess-interview-karthik-vaidyanathan/

Note: Karthik is fairly wrong with his DP4a support listing.
Skylake (Gen9) does not support DP4a and Nvidia actually started to support it with most Pascal GPUs and AMD has in theory the hw capabilities before RDNA2.

Intel showed an illustration in that regard, which is not based on final measured data, but if reality looks close to this, the lack of dedicated matrix units wouldn't be a deal breaker and DP4a acceleration should be enough for a wider range of hardware, not only high-end GPUs.

 

[Locuza]

| Answer Part 2: I always got an error message, without specifying the issue, when I tried to post everything together, the cut was made till the error point |

This list doesn't seem right. Desktop Pascal GPUs don't support FP16 rapid packed math, let alone INT8 or INT4. They even needed to get a single FP16 unit with ridiculously low performance just to get some IEEE compliance.

They can probably still run these algorythms by promoting all variables to FP32 on the driver level, but in that case virtually every DX11 and later GPU could run XeSS (with varying levels of efficiency of course).

Lastly, I'm not sure what the point of this thread is. If XeSS on DP4a is open source then every modern GPU can run it. Series, XBOne S/X, PS4/PS5, Switch, etc.
DP4a is a hardware feature to accelerate calculations using smaller variables, but GPUs can promote these variables up to 16 or 32bit and run the same calculations.

Pascal does support DP4a (INT8) starting with GP102:

https://developer.nvidia.com/blog/mixed-precision-programming-cuda-8/

Some FP16 API operations don't require explicit FP16 support and can be promoted to FP32, which Pascal is doing, however, DP4a with int8 won't be morphed and promoted to FP32.
At launch XeSS will have two backends, one Intel exclusive XMX backend with Matrix Unit support and one cross-vendor backend using the DX12 API and Shader Model 6.4 and 6.6 (if supported).
"Usman: Does XeSS work on hardware with no XMX or DP4a support with an FP16 or FP32 fallback?"
"Karthik: No, not at the moment. We will look into it, but I cannot commit to anything at this point. Even, if you were able to do it, there's the big question of performance and whether its justified."

Perhaps a packed 16-Bit implementation can work reasonably well on some cases, but the advantage with XeSS vs. native rendering will likely be significantly less.

Intel's own Tom Peterson did say in an interview that Intel's cards use their own internal programming language which they say is super optimised for XeSS.

So how well Microsoft's API's for ML can leverage XeSS is open to debate.

I wouldn't expect anything "staggering" but regardless XeSS is still very impressive and I'm curios too see how AMD can respond with their next iteration of FSR.

For XMX they are using their own interfaces, but the DP4a backend uses Microsoft's DX12 PC API.
One great aspect about the DP4a backend will be that AMD hardware will also be supported and since they don't have matrix units yet, there is no strong need to respond with something of their own.
However, relatively soon AMD has to provide a similar solution to stay competitive and in control.

Still makes me chuckle that Intel dropped a solution before AMD, Microsoft and Sony when they don’t yet have a racing horse in all of this.

It’s like some dude who partied all night and seems to not give a care in the world, dropping a perfect score homework on the desk while the rest of the group worked all night with no solutions.

Well, with DG2 they have a horse in this race, it's their first serious consumer dGPU after two decades (1998).
Thankfully Intel is taking it seriously.

 

[Locuza]

| Answer Part 3: I always got an error message, without specifying the issue, when I tried to post everything together |

They mark it as "no" but do have a ? on that line.

Good to see that Xbox series has support. Whether it ever gets used on console or not is anyone's guess.

The questionmark in brackets relates to the PS5.
GFX1013, codename Cyan Skillfish, is the graphics unit of an APU which looks extremely like the PS5.
It uses the older IP version line 10.1, like other Navi1X/RDNA1 GPUs, however, it does support Ray Tracing, something which no other RDNA1 GPU does.
It does not support DP4a instructions, like Navi10, but unlike Navi12/14 which do support DP4a.
Furthermore, it uses the same cache and memory configuration as Navi10, which is like on the PS5 GPU.

In addition, the APU has 8 CPU cores and only 2 Display Controllers, everyone can guess now how many display controllers the PS5 has...., right..., it has 2.
So very likely the PS5 SoC is behind Cyan Skillfish or something directly derived from it.

Summaries from Coalacanth's Dream based on AMD's Linux drivers:
https://www.coelacanth-dream.com/posts/2021/07/21/amd-cyan_skilfish-rdna-apu/
https://www.coelacanth-dream.com/posts/2021/08/01/cyan_skilfish-apu-gfx1013/#gfx1013
https://www.coelacanth-dream.com/posts/2021/09/28/cyan_skilfish-display-support/

[...]
So it seems hardware without ML hardware (int4) will run XeSS but there will be some performance quality/complexity trade off.

XeSS is using INT8, DP4a = dot-product with 4 elements (INT8) accumulate.
DP8a would use INT4.
DP4a support will be required for the cross-vendor backend to work, it won't run without it.
If there will be a 16-Bit backend implementation is currently open.

Quite doubt that PS4 doesn't support int4/int8 calculations, it's not something new that AMD had to implement but a heritage function from RDNA1.

Also Sony was first to introduce(CB) and heavily rely on reconstruction/upscaling features to be so naive about that.

Navi10 (RDNA1) does not support INT4/INT8, GFX1013 isn't either, the latter appears to be the same tech which the PS5 is using.

Great useless response.

Even Vega supports (GCN) it but PS5 not lol

Only Vega20, no other GCN5 part has hw support.

Just to add something here about Navi 10.

It was designed with INT4/INT8 native support but due a bug AMD shipped with these features disabled.

The bug was fixed in post RDNA revisions (Navi 12, Navi 14, etc).

That's an assumption as far as I know and not a fact.
It's also possible that the hardware does not include that featureset at all.

It's pretty obvious such features were likely removed by Sony themselves for several reasons but that's another discussion. The ML hardware for INT4/8 operations has been found in AMD GPU's up to and including RDNA 2, so similar to VRS and SFS if the PS5 doesn't have hardware ML support then it means it was purposefully removed by Sony.

Based on the first GFX IP version for PS5 hardware, 10.0.0, it looks like the PS5 hardware started as the first RDNA GPU and changed over time.
Many features like VRS, DP4a and a couple of others were likely not finished yet/part of the development cycle and not removed at Sony's request.

VRS tier 2 is hardware based VRS, VRS Tier 1 is pure software, PS5 only supports Tier 1, we know that

DX12 VRS Tier 1 is not a pure software solution, it also requires hardware support.
IIRC the presentation slides from the software VRS method in Call of Duty said that some parts of it are comparable to the VRS Tier 1 spec under DX12, this might be the reason why people think VRS 1 is purely a software thing.

[...]
Which is a shame because OP's source material on Locuza 's tweets is pretty good overall, and I doubt he'd ever want it to be used as cannon fodder for console wars.

I knew from the beginning that some console warrior will quote that table and use it to shit on the PS5 platform, but I don't care what console warriors and their emotional investment in a plastic box are doing.
The table is there to provide an overview about the potential hardware reach of XeSS with DP4a, if necessary software support is provided.

I wouldn't be surprised if Sony didn't bother to expose VRS on their API as it's not getting great results out of contained demos from Coalition anyways, and I wonder if it's of any use at all in UE5.

VRS is a VR feature for foveated rendering first and foremost. It could be that Sony only exposes it to PSVR2 devkits.

The ability to control the pixel shading rate is in general quite useful, which is why some games use a software implementation, to further optimize the visual/performance-ratio.
It doesn't appear as if the PS5 has a direct hardware path to control the pixel shading rate.

 

[Epic Sax CEO]

So your saying Cerny made that patent for nothing?

Did you even read how the patent does it?

Did you even read my post that you replied?
I'm not questioning anything that Cerny did there, I'm sure that it must work as he envision. I'm just pointing that when put in practice in this game competition he may had put all that work for nothing. VRS can be a little performance advantage for the SeX that has it over the PS5, it sure should help with it's lower pixel fillrate, right? But when games more to use VRS the same way it it's irrelevant that the PS5 doesn't have VRS now, Cerny modification may also become irrelevant because the SeX will also benefit from the switch.

 

[Sosokrates]

| Answer Part 3: I always got an error message, without specifying the issue, when I tried to post everything together |


The questionmark in brackets relates to the PS5.
GFX1013, codename Cyan Skillfish, is the graphics unit of an APU which looks extremely like the PS5.
It uses the older IP version line 10.1, like other Navi1X/RDNA1 GPUs, however, it does support Ray Tracing, something which no other RDNA1 GPU does.
It does not support DP4a instructions, like Navi10, but unlike Navi12/14 which do support DP4a.
Furthermore, it uses the same cache and memory configuration as Navi10, which is like on the PS5 GPU.

In addition, the APU has 8 CPU cores and only 2 Display Controllers, everyone can guess now how many display controllers the PS5 has...., right..., it has 2.
So very likely the PS5 SoC is behind Cyan Skillfish or something directly derived from it.

Summeries from Coalacanth's Dream based on AMD's Linux drivers:
https://www.coelacanth-dream.com/posts/2021/07/21/amd-cyan_skilfish-rdna-apu/
https://www.coelacanth-dream.com/posts/2021/08/01/cyan_skilfish-apu-gfx1013/#gfx1013
https://www.coelacanth-dream.com/posts/2021/09/28/cyan_skilfish-display-support/


XeSS is using INT8, DP4a = dot-product with 4 elements (INT8) accumulate.
DP8a would use INT4.
DP4a support will be required for the cross-vendor backend to work, it won't run without it.
If there will be a 16-Bit backend implementation is currently open.


Navi10 (RDNA1) does not support INT4/INT8, GFX1013 isn't either, the latter appears to be the same tech which the PS5 is using.


Only Vega20, no other GCN5 part has hw support.


That's an assumption as far as I know and not a fact.
It's also possible that the hardware does not include that featureset at all.


Based on the first GFX IP version for PS5 hardware, 10.0.0, it looks like the PS5 hardware started as the first RDNA GPU and changed over time.
Many features like VRS, DP4a and a couple of others were likely not finished yet/part of the development cycle and not removed at Sony's request.


DX12 VRS Tier 1 is not a pure software solution, it also requires hardware support.
IIRC the presentation slides from the software VRS method in Call of Duty said that some parts of it are comparable to the VRS Tier 1 spec under DX12, this might be the reason why people think VRS 1 is purely a software thing.


I knew from the beginning that some console warrior will quote that table and use it to shit on the PS5 platform, but I don't care what console warriors and their emotional investment in a plastic box are doing.
The table is there to provide an overview about the potential hardware reach of XeSS with DP4a, if necessary software support is provided.


The ability to control the pixel shading rate is in general quite useful, which is why some games use a software implementation, to further optimize the visual/performance-ratio.
It doesn't appear as if the PS5 has a direct hardware path to control the pixel shading rate.

Informative post, thanks for spending the time to type all that out, plus your twitter work.
I have a question, Do you think a 16bit/PS5 implementation of XeSS will be as performant as the current int8 implementation?

 

[Sean Mirrsen]

That's actually an open question.

I don't doubt that it's an open question - there's very little in terms of practical experimentation results out there. But it's a simple matter to theorize about. Without dedicated ML cores, you'd be putting the calculation load on the conventional processing hardware. That's processing hardware that not only has to do more work (more cycles of dot-product calculations compared to matrix-multiplication) for the same result, but is also already under load from the regular rendering and game-running process. So if in an Nvidia card the Tensor cores add on top of the shader cores, and are otherwise used for RT denoising and whatnot, in any other setup that load falls on the shader cores, reducing the available performance for creating the base render. So even if you're running the same logic engine as DLSS (thus ensuring that the output quality is good), you will need much more time to run the reconstruction pass on the shader cores, and that means that in order to get a good framerate the render resolution you'd have DLSS working from would need to further decrease, allowing that extended reconstruction pass to fit in the total frame time.

I don't doubt that if it's allowed to, it'll work - my doubt is in that, unless it's already a higher-end GPU like a 2000 series, there might not really be a net benefit to running DLSS. Anything Nvidia-made stronger than a GTX 1080, already has Tensors. Integrated graphics, generally don't get to be even that powerful. And I doubt Nvidia has much consideration for its direct rival in GPU space.

 

[ToTTenTranz]

I have a question, Do you think a 16bit/PS5 implementation of XeSS will be as performant as the current int8 implementation?

His answer is here:

Perhaps a packed 16-Bit implementation can work reasonably well on some cases, but the advantage with XeSS vs. native rendering will likely be significantly less.

Though a more concrete answer would be: is this combo of GPU + game / load one that particularly taxes the shader processors? If so, then there are gains to be had from using UINT8 at quad-rate instead of FP16 at dual rate.
If the GPU with that load in question isn't getting a great utilization of its shading units then the difference could be small or unnoticeable.

For example, the original FP16 path of FSR is just some 5% faster than the "slow" FP32 one in most RDNA and Vega GPUs. That is not to say FSR is as heavy as XeSS, as it most probably isn't.

 

[ToTTenTranz]

Note: Karthik is fairly wrong with his DP4a support listing.

Perhaps he was talking about Gen11 (Ice Lake) and Gen12 / Xe / DG1 and maybe even Gen10 (a stillborn for all intents and purposes).
To us, Intel went practically from Gen9 to Gen12 with a quick stop on Gen11 that lasted a year and only on 15W laptops. To someone developing these GPUs they went through 2 or 3 whole generations of GPU architectures with DP4a support before reaching Alchemist.

 

[Locuza]

I have a question, Do you think a 16bit/PS5 implementation of XeSS will be as performant as the current int8 implementation?

Based on Intel's comment, it could be outside the range where it's worth it, at least FP32 should be worthless:
"Usman: Does XeSS work on hardware with no XMX or DP4a support with an FP16 or FP32 fallback?"
"Karthik: No, not at the moment. We will look into it, but I cannot commit to anything at this point. Even, if you were able to do it, there's the big question of performance and whether its justified."
https://wccftech.com/intel-xess-interview-karthik-vaidyanathan/

With 16-bit the cost of XeSS could be too high to beat other compromises.
Like upscaling from 1080p to 4K with DP4a may work out, while with packed 16-bit it could be too slow or would need to target another resolution, degrading the visual outcome.
That visual outcome could be worse than simply picking 1440p native res and letting standard TAA do its work.

Most questions can't be confidently answered yet, however just a quarter or so and we will have the answers.

I don't doubt that if it's allowed to, it'll work - my doubt is in that, unless it's already a higher-end GPU like a 2000 series, there might not really be a net benefit to running DLSS. Anything Nvidia-made stronger than a GTX 1080, already has Tensors. Integrated graphics, generally don't get to be even that powerful. And I doubt Nvidia has much consideration for its direct rival in GPU space.

I agree with the thought process behind DLSS and the performance with matrix units vs. running DP4a on Shader Cores, though we will see how it turns out in practise.
Intel's commentary in that regard sounds rather positive.

Perhaps he was talking about Gen11 (Ice Lake) and Gen12 / Xe / DG1 and maybe even Gen10 (a stillborn for all intents and purposes).
To us, Intel went practically from Gen9 to Gen12 with a quick stop on Gen11 that lasted a year and only on 15W laptops. To someone developing these GPUs they went through 2 or 3 whole generations of GPU architectures with DP4a support before reaching Alchemist.

Well, he said Skylake and that older internal GPUs had it for a while now, but that's not true.
DP4a support or "DL Boost" as Intel's marketing liked to call it, came with Xe LP (Gen12.1), Gen 11 under Ice Lake does not support it.

 

[Sosokrates]

Based on Intel's comment, it could be outside the range where it's worth it, at least FP32 should be worthless:
"Usman: Does XeSS work on hardware with no XMX or DP4a support with an FP16 or FP32 fallback?"
"Karthik: No, not at the moment. We will look into it, but I cannot commit to anything at this point. Even, if you were able to do it, there's the big question of performance and whether its justified."
https://wccftech.com/intel-xess-interview-karthik-vaidyanathan/

With 16-bit the cost of XeSS could be too high to beat other compromises.
Like upscaling from 1080p to 4K with DP4a may work out, while with packed 16-bit it could be too slow or would need to target another resolution, degrading the visual outcome.
That visual outcome could be worse than simply picking 1440p native res and letting standard TAA do its work.

Most questions can't be confidently answered yet, however just a quarter or so and we will have the answers.


I agree with the thought process behind DLSS and the performance with matrix units vs. running DP4a on Shader Cores, though we will see how it turns out in practise.
Intel's commentary in that regard sounds rather positive.


Well, he said Skylake and that older internal GPUs had it for a while now, but that's not true.
DP4a support or "DL Boost" as Intel's marketing liked to call it, came with Xe LP (Gen12.1), Gen 11 under Ice Lake does not support it.

I wonder why sony would not include int8 on the PS5 if it helps this much with A.i reconstruction.

 

[Sean Mirrsen]

I agree with the thought process behind DLSS and the performance with matrix units vs. running DP4a on Shader Cores, though we will see how it turns out in practise.
Intel's commentary in that regard sounds rather positive.

My only real positive sort of thought - a 'hope', if you will - is that it's generally been demonstrated that the frame time required for a DLSS pass, is dependent on the output resolution, not the input resolution (for the most part). And we've seen that DLSS can make surprisingly good reconstructed imagery from very low resolutions. So that, if you have a game with a very high render load, and your GPU is some integrated thing (maybe an RDNA2 integrated thing, like the Deck's APU), you could push the resolution way down to get some of that frame time back, and let the slower, shader-core reconstruction process work - for, hopefully, decent image quality at a decent framerate that you wouldn't otherwise get.

 

[Max Payne’s Baretta]

I wonder why sony would not include int8 on the PS5 if it helps this much with A.i reconstruction.

Manufacturers take bets on what technologies will be relevant for their use case. In this situation it’s likely when scoping the console they felt as if the advancement in the sector , thus the applicability, simply wasn’t worth the investment in the silicon.

 

[winjer]

I wonder why sony would not include int8 on the PS5 if it helps this much with A.i reconstruction.

Because Sony wanted to get sooner to manufacture, and couldn't wait for AMD to finish implementing these features.

 

[Arioco]

Because Sony wanted to get sooner to manufacture, and couldn't wait for AMD to finish implementing these features.

Even a Vega card supported it. Maybe SONY couldn't wait for RDNA2 to be finish, but some AMD GPU have been sopporting this feature since before RDNA1. Why wouldn't SONY want to include? It doesn't even make the APU bigger.

 

[Sosokrates]

Manufacturers take bets on what technologies will be relevant for their use case. In this situation it’s likely when scoping the console they felt as if the advancement in the sector , thus the applicability, simply wasn’t worth the investment in the silicon.

This could provide a sizeable advantage for series consoles.

However it will be interesting to see how UE5's implementation compares and of FP16 reconstruction is improved. The PS4 pro used FP16 so theu could just be working on a better version of that.

 

[Max Payne’s Baretta]

Even a Vega card supported it. Maybe SONY couldn't wait for RDNA2 to be finish, but some AMD GPU have been sopporting this feature since before RDNA1. Why wouldn't SONY want to include? It doesn't even make the APU bigger.

Yea that’s true but getting out first assists with being the market leader. Market leader gets lead platform thus resource preference ect ect.

So if these advantages never materialize it doesn’t matter.

 

[Loxus]

The questionmark in brackets relates to the PS5.
GFX1013, codename Cyan Skillfish, is the graphics unit of an APU which looks extremely like the PS5.
It uses the older IP version line 10.1, like other Navi1X/RDNA1 GPUs, however, it does support Ray Tracing, something which no other RDNA1 GPU does.
It does not support DP4a instructions, like Navi10, but unlike Navi12/14 which do support DP4a.
Furthermore, it uses the same cache and memory configuration as Navi10, which is like on the PS5 GPU.

In addition, the APU has 8 CPU cores and only 2 Display Controllers, everyone can guess now how many display controllers the PS5 has...., right..., it has 2.
So very likely the PS5 SoC is behind Cyan Skillfish or something directly derived from it.

Summaries from Coalacanth's Dream based on AMD's Linux drivers:
https://www.coelacanth-dream.com/posts/2021/07/21/amd-cyan_skilfish-rdna-apu/
https://www.coelacanth-dream.com/posts/2021/08/01/cyan_skilfish-apu-gfx1013/#gfx1013
https://www.coelacanth-dream.com/posts/2021/09/28/cyan_skilfish-display-support/

I know you know alot about GPUs, but I have ask what makes you think Cyan Skillfish is the PS5.

The PS5's (Ariel) GPU is already codenamed Oberon with DCN 3.0 and has an RDNA 2 core.

The patches posted for enabling Cyan Skillfish display support, add Display Core Next 2.01 display engine support and is RDNA 1.

Don't you find a bit awkward that the PS5 would seemingly changes it's core to RDNA 1 and step back to DCN 2.01?
It just doesn't make sense that Cyan Skillfish is the PS5.

 

[winjer]

Even a Vega card supported it. Maybe SONY couldn't wait for RDNA2 to be finish, but some AMD GPU have been sopporting this feature since before RDNA1. Why wouldn't SONY want to include? It doesn't even make the APU bigger.

But it had to be implemented in a new arch.
And that takes a bit of time. MS waited a few months. Sony didn't.

 

[Loxus]

But it had to be implemented in a new arch.
And that takes a bit of time. MS waited a few months. Sony didn't.

I don't know why you keep spreading bs.

These chips take years to develop.
Everything that you are going to put into the chip is decided in the early stages. These chips where finalize in 2019. What your saying would of made sense if Microsoft chip was finalized in 2020.

These consoles sometimes go through development issues and can have many revisions.

You think designing a APU is like building with Lego blocks?

 

[MonarchJT]

I don't know why you keep spreading bs.

These chips take years to develop.
Everything that you are going to put into the chip is decided in the early stages. These chips where finalize in 2019. What your saying would of made sense if Microsoft chip was finalized in 2020.

These consoles sometimes go through development issues and can have many revisions.

You think designing a APU is like building with Lego blocks?

the project of the PS5 Apu with test from AMD is much older (at least 1 year) before the the series x.....probably you didn't follow the AMD leak

 

[oldergamer]

I don't know why you keep spreading bs.

These chips take years to develop.
Everything that you are going to put into the chip is decided in the early stages. These chips where finalize in 2019. What your saying would of made sense if Microsoft chip was finalized in 2020.

These consoles sometimes go through development issues and can have many revisions.

You think designing a APU is like building with Lego blocks?

You are still arguing and posting bullshit that is incorrect. Give it a rest misterx.

 

[oldergamer]

the project of the PS5 Apu with test from AMD is much older (at least 1 year) before the the series x.....probably you didn't follow the AMD leak

This is correct!

 

[azertydu91]

You are still arguing and posting bullshit that is incorrect. Give it a rest misterx.

Quick question but wouldn't misterx be defending the Xbox in this case?You know the guy that still think the xbox one has a second gpu hidden in the power brick.

 

[Snake29]

We heard the same misinformation from the xbox fanboys about AMD FFX support, that the PS5 wouldn't support it, and that Sony would be the last to even talk or implement these kind of features. The fact that they didn't talked about it and yet they were the first console that had a game with official support for AMD Fidelity FX.

There is a lot of "PS doesn't talk about this, so they can't have it".

 

[azertydu91]

We heard the same misinformation from the xbox fanboys about AMD FFX support, that the PS5 wouldn't support it, and that Sony would be the last to even talk or implement these kind of features. The fact that they didn't talked about it and yet they were the first console that had a game with official support for AMD Fidelity FX.

There is a lot of "PS doesn't talk about this, so they can't have it".

Now imagine the level of fanboyism to doubt it has RT just after they announced it has....Well that's Battaglia for you a real "expert".

 

[ToTTenTranz]

the project of the PS5 Apu with test from AMD is much older (at least 1 year) before the the series x.....probably you didn't follow the AMD leak

The same github leak that showed no raytracing support on the PS5 and used significantly lower GPU and CPU clocks?

Is there any proof that the 2018 leak of project Oberon wasn't simply using a custom RX 5700 with different GDDR6 chips, together with e.g. a Ryzen 2700 for early development?

That visual outcome could be worse than simply picking 1440p native res and letting standard TAA do its work.

I guess you meant TAAU, but it should be noted that there's no such thing as standard temporal reconstruction, as the implementations have been evolving a lot through time.

For example, UE5's TSR already shows better quality at 2.3x less pixels (see 720p TSR vs. 1080p), whereas an older temporal upsampling used by Bluepoint in Demon's Souls 2020 shows similar but lower quality between native 4K (quality mode) and 2.3x less pixels (1440p + temporal upsampling in performance mode).

The biggest difference between UE5's TSR and Demon's Souls 2020 temporal upscaling is probably the fact that the former is using motion vectors, just like DLSS and XeSS.
According to Epic, TSR is also highly optimized for RDNA2's shader units, which makes it great for Series consoles, the PS5 and Steam Deck. On Steam Deck's small 1280*800 screen we could see its 1.6TFLOPs GPU rendering at only 640*400 and actually see good visual results nonetheless.

Of course, Epic's latest temporal reconstruction doesn't need to be the only one using motion vectors. It looks like Call of Duty Vanguard is already using that approach to get very good results.


Regardless, what I mean to say with this is that temporal reconstruction is seemingly getting really good with great performance, and it seems to be already very competitive with DLSS and XeSS.
If TSR becomes the norm, it could be that AMD doesn't really need to introduce matrix multiplication units for fast DL inference anytime soon for their consumer solutions (CDNA already does have matrix multiply acceleration), and they're better off spending their silicon budget on higher rasterization and raytracing performance instead. It does seems that RDNA3 is again skipping on tensor cores, despite having brutal rasterization performance.

 

[Loxus]

the project of the PS5 Apu with test from AMD is much older (at least 1 year) before the the series x.....probably you didn't follow the AMD leak

You are still arguing and posting bullshit that is incorrect. Give it a rest misterx.

Facts don't lie, the image also says 2019 for Series consoles.
Are you saying I Photoshopped that image?

There is no proof Microsoft waited longer than Sony for anything. Microsoft didn't know shit about PS5 during it development anymore than us trying to believe the so called leakers, everything is under NDA.
Do you even know wtf Microsoft even waited for that isn't in PS5?

An engineering sample is not final silicon.

It's official, you guys are fucked up for wanting a console to fail so badly and call yourselves gamers.

 

[Loxus]

We heard the same misinformation from the xbox fanboys about AMD FFX support, that the PS5 wouldn't support it, and that Sony would be the last to even talk or implement these kind of features. The fact that they didn't talked about it and yet they were the first console that had a game with official support for AMD Fidelity FX.

There is a lot of "PS doesn't talk about this, so they can't have it".

Same for the Unreal Engine 5 demo.
Sony never acknowledge that UE5 demo the way we would expect or even uploaded it to their YouTube channel.