IBM's Ashwini Nanda on Cell blades, raycasting, and more

[one]

Ashwini Nanda leads research, technology and strategy on Cell Processor Based Systems, including the IBM Cell Blades prototype, at IBM TJ Watson Research Center, Yorktown Heights, NY.
http://www.research.ibm.com/people/a/ashwini/

The presentations about Cell blades at E3 including Cell-based Terrain Rendering Engine (TRE), Online game demo, and Video Surveillance Demo (and Cool Chips slides) are uploaded on his page.

http://www.research.ibm.com/people/a/ashwini/E3 2005 Cell Blade reports/

From TRE demo,

TRE Performance 	 

The server has many rendering parameters that effect 	 
performance including: 	 

1) Output image size 	 
2) Map size 	 
3) Visibility to full fog/haze 	 
4) Multi-sampling rate 	 

For benchmarking purposes the following values will 	 
be selected [figure 7]: 	 

1) 1280x720 (720p) output image size 	 
2) 7455x8005 Map size 	 
3) 2048 map steps to full haze (10m map -> 	 
20Km visibility) 	 
4) 1.33 x (2 – 8 Dynamic) or ~2-32 samples 	 
per pixel 	 

With these settings the following rendered and 	 
compressed relative image rates were captured: 	 

Processor  		 	Performance 	 
2.0 GHz G5 VMX 		 		1 (No Image Encode) 	 
2.4 GHz UP Cell 	 		36 	 
3.2 GHz UP Cell 	 		50 	 
2.4 GHz 2-way SMP Cell 		 	75

 

[cho]

so, the "Waternoose" @3.2ghz may reach ~ 5 fps ?

 

[dukmahsik]

and this has to do with ps3 how??

 

[nAo]

and this has to do with ps3 how??

PS3 is CELL based hence all this stuff is very interesting.
RT demo seems to be quite good (well..it's really fast even if it's running at high resolution and with antialiasing..)

 

[SanGreal]

The PS3 should definitely benefit from all of the research other companies are putting into cell.

 

[hasanahmad]

demo demo demo rendering demo! thats all . what about ingame!

 

[one]

demo demo demo rendering demo! thats all . what about ingame!

Aren't you forgetting that there's a front page in Beyond3D and Dave Baumann uses RightMarkD3D to review new vid cards...? Please keep it on topic.

 

[MechanizedDeath]

The output images are great, and would look even better from higher altitude. But the image quality still seems hampered by the granularity of the sat data.

And I wonder if this sort of thing could open up a whole new world for flight sims. I mean the possibility for virtually limitless maps with excellent terrain detail. Maybe not now, but in the future, with any luck. PEACE.

 

[mckmas8808]

I wonder is any of this research going to be implemented in the PS3 for game purposes. Do you guys think that any of this stuff can happen.

 

[Farid]

Thanks for the heads up, one.

demo demo demo rendering demo! thats all . what about ingame!

This was uncalled for, you know? There's a time for everything, in the 3D timeline, first come the tech demos and then come the games.
And since we're in a 3D technical forum, tech demo discussions have their place here, as much as actual game engines discussions.

 

[Carl B]

@One: Great find! Not too much new information, but it's nice being able to read over the white-papers. The ray-casting information is especially interesting, though I hope we're able to see something more in-depth one day in the not to distant future; both demo-wise and benchmark-wise. The resolution and framerate achieved was impressive though, and the fact that a 2-way system showed a straight 2x increase in performance bodes well for customer like Mercury Systems, who is now going the route of ray-tracing solutions for some customers.

 

[Guden Oden]

demo demo demo rendering demo! thats all . what about ingame!

I wouldn't call this a demo. It's certainly not a game demo in the traditional sense, rather it's a research application. Terrain rendering has many applications in games however and as such I'm sure this experiment will be very useful - even though I don't expect terrain rendering to be done on the host CPU itself.

Besides, more examples that utilize cell and its capabilities well will help programmers to write better game code for the thing as well...

 

[Titanio]

Processor Performance
2.0 GHz G5 VMX 1 (No Image Encode)
2.4 GHz UP Cell 36
3.2 GHz UP Cell 50
2.4 GHz 2-way SMP Cell 75

STI's "10x the performance of conventional processors" claim..conservative?

Nice find one, thanks.

 

[PC-Engine]

TRE? Will CELL being doing this for PS3 games? How about some Linpack benchmarks?

 

[PeterT]

Interesting. This is an application where Cell is probably at least twice as fast as the 360's triple-core cpu, and where it shows an extreme advantage compared to conventional general purpose cpus. Really makes me want to have one, and a way to run my own code on it. If Sony provides that (and I very doubt it, sadly), PS3 will be the first console I'll ever buy at launch, regardless of price.

 

[Titanio]

Interesting. This is an application where Cell is probably at least twice as fast as the 360's triple-core cpu, and where it shows an extreme advantage compared to conventional general purpose cpus. Really makes me want to have one, and a way to run my own code on it. If Sony provides that (and I very doubt it, sadly), PS3 will be the first console I'll ever buy at launch, regardless of price.

It's been mooted that Linux may be pre-installed on all PS3 HDDs, and Kutaragi has talked quite a bit about the need to nurture software development on Cell through an open community. So I'm quite hopeful that'll be possible.

 

[one]

STI's "10x the performance of conventional processors" claim..conservative?

Hehe, also the performance number 50 for 3.2 Ghz uniprocessor Cell is effectively taken with 7 SPEs as 1 SPE is used for the JPEG image compression kernel, so it seems 50 is the number for the PS3 Cell too.

 

[Guden Oden]

I wonder how the SPEs are employed in this coding example, if all of them do the same work, or if data is shuffled around between them in an assembly line fashion, and if so, in how many steps...

 

[MfA]

Wow, someone else read the vertical ray coherence paper ... Ive never understood why that one isn't quoted more than it is, basically the only right way to raycast a heightfield.

You wouldnt want to use this for software rendering in the PS3, but you could probably adapt this method to do conservative occlusion culling, and then actually render the heightfield with polygons as normal ... so it might not be completely without practical use.

Guden, you could just read the paper Basically it just renders "scanlines" in parallel.

 

[ralexand]

PS3 is CELL based hence all this stuff is very interesting.
RT demo seems to be quite good (well..it's really fast even if it's running at high resolution and with antialiasing..)

Was this being output in realtime. I thought it was a series of jpeg assembled and displayed. I have no idea how this compares to what they could do realtime. To be honest I don't see any significant qualitative difference with the images over what could be done with your standard shaders.