nVidia shader patent (REYES/Raytracing/GI) destined for PS3?

[j^aws]

How many possibilities are that NVidia plans one GPU for PS3 with one 'shader Core'(vertex+pixel)?
For example 48 ALU in 2x24 pipeline or 3x16 pipeline.

If the rumoured R500 Xenon GPU will have 48 unified ALUs from the leaked specs and is suspected to be manufactured on 90nm process, then I would expect the nVidia GPU to have 64 ALUs on a 65nm process and with a good dose of eDRAM. ...And possibly 2x32 pipeline configurations...though I couldn't tell you the benefits of 1x64, 2x32, 4x16 for what they would want to achieve with CELL, any guesses?

I would think these ALUs would be unified cores that would be more fragment/pixel oriented as the CELLs cores already have units suitable for vertex processing.

 

[j^aws]

Could it be that next gen differentiator be that Cell architeture would allow ray tracing/Reyes and Xenon next system a more traditional approach to shadeing yet with much more performance than todays gpu, per pixel lighting, etc...?

AFAIK, there will be low level access granted to devs on both systems and both architectures should be flexible enough to try these things. The difference seems to be both the PS3 CPU (CELL) and the nVidia GPU will be stream processors but only the XeGPU will be a stream processor. What this means to developers, I couldn't tell you but they'll sure have fun finding out!

From what we know of guestimates from Xenon leaked diagrams and recent CELL announcements,

Xe CPU @ 3.5 GHz ~ 84 Gflops
PS3 CPU @ 4.6 GHz ~ 294 Gflops

PS3 CPU 3.5x > Xe CPU

And if Xe GPU is on 90 nm process and PS3 GPU is on 65 nm process you would expect a doubling in transistor counts for a given die size on 65nm. So from Xenon leak,

Xe GPU = 240 GFlops

PS3 GPU 2x > Xe GPU

So as a very rough scaling calculation,

PS3 (CPU=294 + GPU=480) 2.4x > Xe (CPU=84 + GPU=240)

However, this is just a guestimate on peak performance and there are may other factors innvolved. Also Xe could have more cores or PS3 more CELLs/ PEs etc by the time of release. And no flames please, it's just a rough guestimate!

If this is any indication then the extra power could translate into effectively trying new things as REYES/raytracing etc. but what I'm most interested in is how that would translate into gameplay and image quality on screen for games! The only thing I have to go on is the ~100 Gflops Sony GScube (+ SGI origin onyx) and if you we're impressed by that in 2001, then sure as heck, you'd be impressed by these consoles next gen for ~ $300!

 

[j^aws]

Raytracing on GPUs seem like a dead end to me.

The benefit of raytracing is in the way it allows a more correct handling of refraction and specular reflection. This makes it ill suited for a stream based approach because it introduce a data dependent step that has, potentially, highly irregular access patterns (cache busting).

I can see GPUs as accelerating first intersection rays (which is also what the above paper deals with), but these account for just a fraction of the work a heavy duty raytracer has to do.

Cheers
Gubbi

That paper mentions that by using tile rendering you can use the cache more efficiently. Perhaps it's time for nVidia to introduce Gigapixels TBDR technology into the GPU 8)

 

[Gubbi]

It still only deals with first intersection rays.

The problem is exactly the same for CELL; poor random memory access.

Cheers
Gubbi

 

[Megadrive1988]

AFAIK, there will be low level access granted to devs on both systems and both architectures should be flexible enough to try these things. The difference seems to be both the PS3 CPU (CELL) and the nVidia GPU will be stream processors but only the XeGPU will be a stream processor. What this means to developers, I couldn't tell you but they'll sure have fun finding out! Wink

From what we know of guestimates from Xenon leaked diagrams and recent CELL announcements,

Xe CPU @ 3.5 GHz ~ 84 Gflops
PS3 CPU @ 4.6 GHz ~ 294 Gflops

PS3 CPU 3.5x > Xe CPU

And if Xe GPU is on 90 nm process and PS3 GPU is on 65 nm process you would expect a doubling in transistor counts for a given die size on 65nm. So,

PS3 GPU 2x > Xe GPU

So as a very rough scaling calculation,

PS3 (CPU + GPU) 7x > Xe (CPU + GPU)

However, this is just a guestimate on peak performance and there are may other factors innvolved. Also Xe could have more cores or PS3 more CELLs/ PEs etc by the time of release. And no flames please, it's just a rough guestimate! Wink

If this is any indication then the extra power could translate into effectively trying new things as REYES/raytracing etc. but what I'm most interested in is how that would translate into gameplay and image quality on screen for games! Razz The only thing I have to go on is the ~100 Gflops Sony GScube (+ SGI origin onyx) and if you we're impressed by that in 2001, then sure as heck, you'd be impressed by these consoles next gen for ~ $300! Wink

well, all we can say with near-certainty right now, is that Xe CPU will likely be sub 100 gflops or at most, around 100 Gflops. (maybe they've added a forth core or its always been a 4-core CPU as some reports indicated, i'll dig them up after work)

PS3 CPU is likely to be in the 100s of Gflops peak performance. whether it's the low 100s or upper 100s, we don't know. the whole 1 Tflop thing is so unimportant at this stage, i don't even want to think about that anymore. it's enough to say that PS3's successor in the early part of the next decade, will be a multi Tflops machine.

so it's looking likely that PS3 CPU will outclass Xe CPU by several times in peak performance. it will all come down to what the VPU/GPU in Xe and PS3 can output to the screen anyway, and there, things are likely to be more equal (ATI-MS vs Nvidia-Sony), that is, if the PS3 GPU is heavily based on nextgen GeForce architecture. I could well be wrong, and Vince suggests that those thinking that PS3 GPU will be like typical PC GPU architecture are dead wrong. it could well be that Sony is cooking up a 'GeForceSynthesizer' or 'Cell-Force' GPU, with ALL of the strong points of the GS family (GS, GS I-32, GS2, GS3/Visualizer) plus all of the strong points of Nvidia's nextgen GPU.


btw, the GSCube came out in 2000, iirc.

 

[Mythos]

Thanx Jaws.

The only thing I have to go on is the ~100 Gflops Sony GScube (+ SGI origin onyx) and if you we're impressed by that in 2001, then sure as heck, you'd be impressed by these consoles next gen for ~ $300!

I remeber the GScube presentation in 2001 @ SIGGRAPH. However, given some of the comments Sony stated about the GScube (they wanted more performance) ain't it more likely that with Cell/Nvidia that performance targets are around GScube 64 (realistic rendering in realtime).[/quote]

 

[j^aws]

Megadrive,

BTW, I've edited my above post slightly for one CELL, in GFlops,

PS3 (CPU=294 + GPU=480) 2.4x > Xe (CPU=84 + GPU=240)

Btw, it was both in 2000 and 2001, iirc. The latter had Square render FFSW also on nVidia hardware. Both Sony's and nVidia strategy/R&D has been to bridge the gap between offline rendering and realtime rendering. And this patent on Exluna rendering tech would be neat for a hardware implementation. i think both nVidia's and Sonys goals could be met with the PS3. Perhaps nVidia was restricted by MS/DX10/DXNext and Sony complememted it's future direction better?

Mythos,

If GScube 64 ~ 400 GFlops (not sure on SGI but they had 32 Mips R1400 cores, iirc), so 600-800 GFlop prolly in total (I'll add Mips if I find the figures...), then as in above, one CELL + GPU ~ 800 GFlops. So it would seem to fall in the same ballpark as a GScube-64 but it would have way less memory.

 

[j^aws]

...
so it's looking likely that PS3 CPU will outclass Xe CPU by several times in peak performance. it will all come down to what the VPU/GPU in Xe and PS3 can output to the screen anyway, and there, things are likely to be more equal (ATI-MS vs Nvidia-Sony), that is, if the PS3 GPU is heavily based on nextgen GeForce architecture. I could well be wrong, and Vince suggests that those thinking that PS3 GPU will be like typical PC GPU architecture are dead wrong. it could well be that Sony is cooking up a 'GeForceSynthesizer' or 'Cell-Force' GPU, with ALL of the strong points of the GS family (GS, GS I-32, GS2, GS3/Visualizer) plus all of the strong points of Nvidia's nextgen GPU.

Performance between Xe GPU and PS3 GPU will largely depend on the Process used and die size (same would apply to CPUs, although CELL has an inherent advantage from a clean design). If equal in die size but Xe GPU is on 90nm and PS3 GPU is 65nm, roughly scaling, you should be able to get twice as many transistors at 65nm that would roughly equate to twice the performance.

My current thinking is that from Hofstee's presentation,

CPU<=>GPU--->output

This CPU|GPU logic can be on one IC and can scale to many ICs or you can split the CPU and GPU portions to separate ICs and scale that way (i.e PS3). We should see SLI, distributed rendering tech somewhere inherently built into the GPU to complement the ditributed processing of CELL. This would also lead me to beleive that the GPU would be CELL based and run software Cells.

I would think that the GPU would be some sort of unified ALUs but more focused for fragment/pixel processing as CELL is already a vertex powerhouse. And maybe you could load balance software Cells between CPU and GPU. In fact if you think of a current GPU with vertex and fragment processors but sclae it up, it should resemble the entire PS3 CPU(vertex)+GPU(fragment) chipset! However, these components in the GPU could be mixed Sony-Nvidia IPs, e.g Salc/Salps or pixel engines etc.

Also a good chance to see Gigapixel, TBDR to complement the parallel brick/tile rendering SCE patent!

We should see some Gelato rendering tech from this patent by Larry Gritz somewhere with REYES and raytracing/global illumination enabled shaders. 8)

And maybe finally they may roll out the Voodoo brand name...Voodoo Cells anyone!

 

[Fox5]

http://online.cs.nps.navy.mil/DistanceEducation/online.siggraph.org/2002/Papers/13_GraphicsHardware/purcell.ppt

It works for me...it's a direct link...PM me if it still doesn't work...

I got it here if anyone wants

http://cggmwww.csie.nctu.edu.tw/sem...tracing.on.programmable.graphics.hardware.ppt

Interesting R8500 is better than GF3, I am right?

It says using limited fixed function though...the R8500 was able to use its fixed function abilities along with its dx8.1 abilities.

 

[Megadrive1988]

that's alot of stuff to think about Jaws. I don't really disagree, but then I don't really know either.


being slightly pessimistic (and cautious)
I believe that Sony will at least reach the goals it had for PS2, in the PS3. that is, realtime PS1 CG in games. maybe even somewhat beyond PS1 CG but I should not say one way or another since I don't know.

and then Sony should be able reach the goals it had for PS3, in the PS4. i.e. multi Tflop performance + raytracing + global illumination + full hardware accelerated software rendering, 1000 or 1000s of times more performance than PS2. etc.


Although I am now counting on a very very 'solid' PS3. maybe not with as much performance as we once thought or wanted (yes PS3 will be very powerful) yet, with much much greater visual quality (Nvidia) and relative ease of development (better than even PSP) for developers who just want to get started. PS3 will be better balanced overall than I once thought (before Nvidia was there).

yes it will be difficult to make full use of the PS3 CPU for some years, but it will get there. the situation will be much better than the early days of PS2, and even now with PS2. although this says *nothing* about the amount of resources and manpower needed to make stunning large scale nextgen games. that will still be emense.

DaveB, Vince and Pana, do you guys disagree with what I'm saying?

 

[Mythos]

A little off the subject...

But IF Cell/GPU reach Upper level Flops with Ray tracing, GI, etc..We may see a new cross pollunation with a version of a game containing a short film (or other forms of interactivity). War devil has mention this, movement of EA to LA studios.

 

[one]

and then Sony should be able reach the goals it had for PS3, in the PS4. i.e. multi Tflop performance + raytracing + global illumination + full hardware accelerated software rendering, 1000 or 1000s of times more performance than PS2. etc.

Before PS4, let's make an SLI-ed PS3 8) A PS3 should be with an optical interface, so...

 

[Megadrive1988]

sure, i would love SLI PS3s. stack a few, or many PS3s together. combined their processing power (not over the internet).

 

[DeanoC]

Performance between Xe GPU and PS3 GPU will largely depend on the Process used and die size (same would apply to CPUs, although CELL has an inherent advantage from a clean design). If equal in die size but Xe GPU is on 90nm and PS3 GPU is 65nm, roughly scaling, you should be able to get twice as many transistors at 65nm that would roughly equate to twice the performance.

First, how do you know that Xenon CPU isn't a clean design?

Secondly, how do you know there is a process advantage?

CPU<=>GPU--->output

This applies equally to Xenon as much as PS3, Xenon has been built with good connections between CPU and GPU.

 

[AbbA]

This CPU|GPU logic can be on one IC and can scale to many ICs or you can split the CPU and GPU portions to separate ICs and scale that way (i.e PS3). We should see SLI, distributed rendering tech somewhere inherently built into the GPU to complement the ditributed processing of CELL. This would also lead me to beleive that the GPU would be CELL based and run software Cells.

I would think that the GPU would be some sort of unified ALUs but more focused for fragment/pixel processing as CELL is already a vertex powerhouse. And maybe you could load balance software Cells between CPU and GPU. In fact if you think of a current GPU with vertex and fragment processors but sclae it up, it should resemble the entire PS3 CPU(vertex)+GPU(fragment) chipset! However, these components in the GPU could be mixed Sony-Nvidia IPs, e.g Salc/Salps or pixel engines etc.

Also a good chance to see Gigapixel, TBDR to complement the parallel brick/tile rendering SCE patent!

We should see some Gelato rendering tech from this patent by Larry Gritz somewhere with REYES and raytracing/global illumination enabled shaders. 8)

And maybe finally they may roll out the Voodoo brand name...Voodoo Cells anyone!

Is it usefull to have a "next generation" Hardware with a geometric stage separated to rasterization phase? Vertex in the Cell CPU and fragment/pixel in the NVidia GPU?

What kind of connection between Cell and GPU, and i think it would not be best to have a unified 'shader core' in a GPU to perform vertex+pixel operatrion?

 

[version]

This CPU|GPU logic can be on one IC and can scale to many ICs or you can split the CPU and GPU portions to separate ICs and scale that way (i.e PS3). We should see SLI, distributed rendering tech somewhere inherently built into the GPU to complement the ditributed processing of CELL. This would also lead me to beleive that the GPU would be CELL based and run software Cells.

I would think that the GPU would be some sort of unified ALUs but more focused for fragment/pixel processing as CELL is already a vertex powerhouse. And maybe you could load balance software Cells between CPU and GPU. In fact if you think of a current GPU with vertex and fragment processors but sclae it up, it should resemble the entire PS3 CPU(vertex)+GPU(fragment) chipset! However, these components in the GPU could be mixed Sony-Nvidia IPs, e.g Salc/Salps or pixel engines etc.

Also a good chance to see Gigapixel, TBDR to complement the parallel brick/tile rendering SCE patent!

We should see some Gelato rendering tech from this patent by Larry Gritz somewhere with REYES and raytracing/global illumination enabled shaders. 8)

And maybe finally they may roll out the Voodoo brand name...Voodoo Cells anyone!

Is it usefull to have a "next generation" Hardware with a geometric stage separated to rasterization phase? Vertex in the Cell CPU and fragment/pixel in the NVidia GPU?

What kind of connection between Cell and GPU, and i think it would not be best to have a unified 'shader core' in a GPU to perform vertex+pixel operatrion?

between CPU and GPU is REDWOOD connection it is fast 25-100 GB/s,but texture upload too slow if polygon datas send CPU for GPU

nvidia say ps3 performance 50 *ps2= 3 gigapolygon/s
1 vertex about 16 byte(xyz,color,uv), 3 Ggpoly/s=50 GB/s

if REDWOOD 100 GB/s then texture upload is fine , otherwisw is shit

 

[London Geezer]

Now, this might be an itsy pitsy little OT, but i have a question for you people.

If the PS3 GPU is based on NVIDIA's next gen architecture (NV50), why would NVIDIA cancel the NV50 for PC? Doesn't make sense.

 

[version]

Now, this might be an itsy pitsy little OT, but i have a question for you people.

If the PS3 GPU is based on NVIDIA's next gen architecture (NV50), why would NVIDIA cancel the NV50 for PC? Doesn't make sense.

1. new type of pixelpipeline?
2. embended memory is very fast , probably it ?
3. raytracing? not believe 10 teraflops the minimum for it
3. go out pc biznis ?

 

[London Geezer]

The thing is, i don't see why they would cancel their PC product if their PS3 GPU will be based on the same architecture!!
Of course it's just a rumour, but even as a rumour it sounds strange.
It's not like they haven't got the capabilities to work on more than one chip at the same time... Especially if the 2 chips are similar...?

 

[version]

The thing is, i don't see why they would cancel their PC product if their PS3 GPU will be based on the same architecture!!
Of course it's just a rumour, but even as a rumour it sounds strange.
It's not like they haven't got the capabilities to work on more than one chip at the same time... Especially if the 2 chips are similar...?

pc will run ps3's games ?, NO, sony not idiot