"Cell Technology for Graphics and Visualization"

[cho]

A great paper about Cell on graphics appear on grapicshardwre.org:

http://www.graphicshardware.org/program.html

some slides in that paper:

The Cell and GPUs
Extension to shader pipeline on GPUs
− Vertex shaders for geometric modeling
− NURBS
− Subdivision surfaces
− Continuous level of detail
− Vertex shaders for a additional lighting models
Physically Based Modeling
− Soft-body dynamics
− Rigid-body dynamics
Image Processing
− Background Subtraction for Video Surveillance
Global Illumination
– Raytracing
− Raycasting Terrains
− Rendering of DEMs
− Surface shaders for lighting, shadows

Multi-resolution Subdivision

30+ frames per second with only one Cell processor
– No graphics adapter assist
– 1280x720 (HD 720P) resolution
..
-Written completely in C with intrinsics
..

 

[Shifty Geezer]

Thanks for the link. Very interesting reading with some insight to designing for Cell. Of note was the raytrace demo particulars...

30+ FPS 1280x720 with no GPU assist
Includes...

• Ray/Terrain intersection computation
• Texture Filtering
• Normal computation
• Bump map computation
• Diffuse+Ambient Lighting
• Perlin Noise Clouds
• Atmosphere effects (haze, sun, halo)
• Dynamic multisampling (4-16 samples per pixel)
• Image based input (16 bit height + 16-bit texture
• 29 KB SPE object code
• 224 KB SPE data

MJPEG compression via SPE

That's a 720p output at 30fps with multisampling FSAA! That's pretty darned impressive IMO.

 

[Carl B]

I haven't been able to pull up the entire slideshow as yet (due to my bugged up Acrobat at work), but it looks as if there are some insights here in the slides Cho posted into how Cell would work in conjunction with a GPU (RSX?).

 

[Shifty Geezer]

The show a situation where Cell generates vertex and texture info to feed a GPU. There's no two-way communication shown so the slide seem more centred on Cell+Generic GPU rather than Cell+RSX in PS3.

Tell a lie. The diagram shows a two headed arrow. But there's no talk of GPU feeding Cell data, so it's not useful for looking at RSX's relationship with Cell.

Note : When viewed in a web browser the graphics section crashed out. I had to view the show in standalone Adobe Reader, and even then a couple of slides take an AGE to show.

 

[Carl B]

I finally was able to view the entire slide show by scrolling very slowly through it in my browser. Quite impressive! Yeah you're right that it doesn't seem to indicate the model that RSX+Cell would represent, but Cell's abilities alone are quite exciting there. And the slides with a 'GPU' included in the model do also seem to imply Cell's ability to help with geometry and environments.

 

[Shifty Geezer]

Also makes one wonder how well RSX+Cell could perform in raytracing? If Cell alone could manage as much as it did, if model processing and texturing could be handled by RSX, that should free up Cell to maintain higher frame-rates or help with subdivision surfaces. Would Cell+RSX manage the demo at 1080p? And for the PS3's games where a full raytraced scene would be pointless, this does certainly suggest to me the liklihood of true traced reflections on some objects, and even the crazy chance of full raytraced puzzlers. Tetris has never looked so good!

But also it shows there isn't enough oomph for any degree of traced GI. Though maybe a degree of hack could be used that perhaps blends between different prebaked GI maps or something?

 

[pixelbox]

so it take 2 spe's for physics....

 

[Shifty Geezer]

As this was a demo it's no use for games. There'd be nothing left for the actual game. But I'm sure a degree of tracing using some 5-10% could be comfortably included, such as 30fps true reflections in GT5.

 

[pixelbox]

As this was a demo it's no use for games. There'd be nothing left for the actual game. But I'm sure a degree of tracing using some 5-10% could be comfortably included, such as 30fps true reflections in GT5.

not only that but ps3 has 7 spe's. i dont know what a missing spe could do to the performance. nice read though

 

[Carl B]

not only that but ps3 has 7 spe's. i dont know what a missing spe could do to the performance. nice read though

I don't think it would be much.

That was just an arbitrary breakdown of SPE uses, afterall - not a represenation based on any actual game code.

 

[Shifty Geezer]

i dont know what a missing spe could do to the performance.

Drop performance by 1/8th . They also don't list clock-speed. Wasn't that mentioned elsewhere? Weren't the blade server CPUs clocked lower than PS3s? Yes, a quick Google shows the Blades are running at up to 2.8 GHz (http://www.theregister.co.uk/2005/05/27/ibm_demos_cell/) which levels out performance somewhat (all other things being equal)

 

[pixelbox]

which levels out performance somewhat

thanks i was on another forum and they were saying that cell couldn't do raytracing without the 8th spe, guess they didnt do enough research

 

[Titanio]

That was just an arbitrary breakdown of SPE uses, afterall - not a represenation based on any actual game code.

I guess it's at least a "reasonable" model that they're proposing there. Although as always, this is where not having the actual presentation to watch can make things tricky i.e. we don't know exactly what they were saying about that slide.

That said, a PPE alone on game code could be reasonable for some games So a PPE + a couple for AI, and others working on tasks that are more ambiguous in terms of what's "game" and what's graphics (physically based modelling) could be more balanced in that regard even.

I wouldn't be at all surprised if some very graphics-orientated, PS3-orientated, dev heavily bias their Cell use toward things that help graphics (or the "movement" of graphics, physical modelling etc.).

 

[Shifty Geezer]

That eighth SPE isn't the key force behind it all. It contributes, but they could raytrace one 4 SPEs or even 1, just not as fast. With no AA they could get away with 4 SPEs, 720p, 30 fps accoridng these IBMs details. Plus I think one SPE was taken up with MJPEG encoding. Not sure if that was it's whole application or if it multithreaded encoding and raytracing. That'd also be a requirement not needed in PS3.

One the one hand I think this'd be a cool demo to ship with the console. On the other I doubt the mainstream would appreciate it as vast terrain can be rendered already with other techniques. But for me, seeing that running realtime would send a shivver up my spine.

 

[scificube]

Thanks for this awesome find! It's going to take a while to sift through all this stuff...what I can at least.

 

[two]

http://www.graphicshardware.org/presentations/bando-hexagonal-gh05.pdf

Any relation with PS3?
GS2?

 

[Carl B]

I guess it's at least a "reasonable" model that they're proposing there. Although as always, this is where not having the actual presentation to watch can make things tricky i.e. we don't know exactly what they were saying about that slide.

That said, a PPE alone on game code could be reasonable for some games So a PPE + a couple for AI, and others working on tasks that are more ambiguous in terms of what's "game" and what's graphics (physically based modelling) could be more balanced in that regard even.

I wouldn't be at all surprised if some very graphics-orientated, PS3-orientated, dev heavily bias their Cell use toward things that help graphics (or the "movement" of graphics, physical modelling etc.).

I was more just addressing the question posed "i dont know what a missing spe could do to the performance..."

Just trying to convey that a missing SPE doesn't mean one of those 'boxes' given in the slide suddenly becomes unworkable, like the 'physics' box in the slide. Shift answered best though with his simple 'all it means is 1/8 less the power...' I agree totally though that SPE/computational distribution will probably vary wildly depending on the game genre. I like what Cell's bringing to the table here.

 

[NAP]

Hi everybody I'm a new customer and I'm Italian so you apologize me if I make some mistake.

To coming IT, if i don't mistake one SPE of cell processor generate 800 million of polygon per second with disabled effects and lights, so if the cell is used for graphics only we have 6.4 billion of polygon per second that's a superlative result, however we know that the cell it's used for other processings like physics, music and A.I. , but it's however rest a good result


Sorry for my mistakes

 

[Titanio]

WRT the terrain demo, it's worth noting in terms of "freeing up" SPEs for other things in a scenario where Cell was working alone, a 7-SPE Cell @ 3.2Ghz (PS3's Cell) could afford to free one of the SPEs from raycasting/rendering if rendering locally, and still maintain that framerate achieved at 2.8Ghz (if everything scales nicely in terms of clockspeed and number of SPEs).

 

[Nite_Hawk]

Hi everybody I'm a new customer and I'm Italian so you apologize me if I make some mistake.

To coming IT, if i don't mistake one SPE of cell processor generate 800 million of polygon per second with disabled effects and lights, so if the cell is used for graphics only we have 6.4 billion of polygon per second that's a superlative result, however we know that the cell it's used for other processings like physics, music and A.I. , but it's however rest a good result


Sorry for my mistakes

Hi NAP,

Welcome to the board. I think we all were expecting Cell to be good at doing physics and possibly AI. Seeing graphics demos on it like this is just amazing in my opinion though.

Nite_Hawk