A software renderer on CELL by IBM

[Titanio]

There's a new video of this tech running across 3 PS3s here:

http://www.youtube.com/watch?v=oLte5f34ya8

http://gametomorrow.com/blog/index.php/2007/04/03/ps3-clusters-2/

 

[Rainbow Man]

The overwhelming majority of the comments to that youtube video are immensely depressing and makes me want to just shove humanity as a whole over the edge of a cliff.

Cool video though and interesting to know it doe snot need any cell blades at all. Good find.
Peace.

 

[Shifty Geezer]

That is way cool seeng the networked PS3s...but also very pointless. The car is no better looking than a realtime GPU render. The link says the model is 75x more complex than the models used in games. I'd have to ask why? That's too much detail. What happens when you move the triangles out of the car and into the scenery? How does the engine perform then.

I'm sure some super-geeks will have a field-day with networked PS3 computing. When these things are $200, some people will populate their room with a little army of number-crunchers for the fun of it. It's very good to see that as an option, with Cell cooperation. But they really need to get some good demos out there! Rendering cars a little less realistic looking than GTHD isn't going to wow many folks or prove any worth to their software renderer!

 

[Titanio]

If the car model is the same as the one in this paper (i think it is?), without the ambient occlusion, then I'd say it's probably better looking than GT:HD

It is probably overkill in terms of geometric complexity, but I don't think the point was to be artistically/visually efficient about it. I guess the point of this demo was how quickly they managed to network some PS3s together toward a task.

I would like to see what a single PS3 can do, and as you suggest, with that geometric complexity spread out over a scene (the urban environment in the OP might be suitable in that regard - 3m polys in that scene, but obviously a spatially much larger model). If you can get the same performance with the same number of polys composed arbitrarily, then I'm sure there are artists who would make better use of that budget than ploughing them all into a single car model

 

[Shifty Geezer]

They need to grab an existing game scene, maybe 2 million polys - okay, 1 million as there won't be shadow volumes to compute - and throw that out there to show the engine can handle discrete objects effectively. The other important thing I think is to showcase RT being worthwhile, unless it's just an academic exercise (which I think it is). The article said it was using a 'rendering algorithm only deployed by the film industry' which is misleading. They don't use it where they don't have to!

The .pdf was more interesting with it's fast AO calculation. That could be worthwhile.

 

[Rainbow Man]

I wonder how many PS3s can be networked before the network itself and perhaps more importantly that single SPE dedicated to decompression and compositing becomes a bottleneck..
Peace.

 

[sunscar]

Right now, Terra Soft produces 32 node PS3 systems, so there seems to be some wiggle room in regard to the network.

And I believe it's a single SPE per PS3 for compression, not a single SPE in general, so that shouldn't be that great an issue.

 

[Megadrive1988]

if I understand this right, only the CELL CPUs of the PS3's are being used, correct?


I wonder if Sony or anyone are going to network some PS3's so that both the CELLs and the RSXs can be used to render images, like the old GSCube.

I mean it's interesting what can be done with some networked CELLs, but imagine if the full power of each PS3 could be used, not just the CELLs.

 

[Shifty Geezer]

As they're Linux Cell demos, compatigle with Linux Blades etc., there's no reason for IBM to want to include RSX in the rendering process even if they could. RSX is meaningless for the machines they'll be selling. If there's to be an open hardware networked processing system, it needs Sony to create it.

 

[Rainbow Man]

Right now, Terra Soft produces 32 node PS3 systems, so there seems to be some wiggle room in regard to the network.

Well that some company's making 32 node systems doesn't guarantee that this application doesn't bottleneck the network (much) before 32 nodes is reached.

But considering the framerate and that that information is compressed I suppose not much network capacity was used by that 3-machine demo. So depending on how much they can compress a pretty nice framerate should be achievable. Gigabit ethernet can get 100+ MB/s transfer rates from what I understand.

And I believe it's a single SPE per PS3 for compression, not a single SPE in general, so that shouldn't be that great an issue.

I said DEcompression. As in a single SPE in the destination machine. The one that produces th efinal monitor output image.

Peace.

 

[inefficient]

That is very very cool.

Generating occlusion maps in real time is quite exciting.

 

[Shifty Geezer]

I said DEcompression. As in a single SPE in the destination machine. The one that produces th efinal monitor output image.

All it's decompressing are final tiles. There'll be enough tiles to make the final image - from one 1280x720 tile to 10 128x720 tiles to 720p broken into a thousand tiles. Unless there's a massive overhead in the decompression per tile, it shouldn't be an issue.

If the scene can fit into each PS3 node's internal RAM, network traffic will just be image tiles and a bit of control code telling the nodes what bit of the scene to render.

 

[Rainbow Man]

All it's decompressing are final tiles. There'll be enough tiles to make the final image - from one 1280x720 tile to 10 128x720 tiles to 720p broken into a thousand tiles.

Yes I understand this. But if there are a lot of PS3s calculating tiles the total swarm of machines should produce lots of tiles meaning a high framerate. So there would be more and more tiles to decompress (per time unit) the more machines are added.

Unless there's a massive overhead in the decompression per tile, it shouldn't be an issue.

Well not knowi g what the specifics are that may or may not be accurate.

But I guess this is more of a proof of concept than an actual product so whenever something bottlenecks it won't be that big a deal anyway hehe.



Peace.

 

[Shifty Geezer]

Yes I understand this. But if there are a lot of PS3s calculating tiles the total swarm of machines should produce lots of tiles meaning a high framerate. So there would be more and more tiles to decompress (per time unit) the more machines are added.

If you mean rendering far faster than 60 frames per second, yes that'd be an issue, similar to decompressing an HD movie at very high framerates. That's going to be a problem with any system! Presumably though in networked render system like this, you'd be using the processing power to add more detail rather than render faster. If you're network is capable of rendering 200 fps and the final Cell can't decode the tiles fast enough, you ought to be rendering more complex data! Or alternative, you just save al the tiles out and when it comes time to play, decode them at the final 60 fps framerate.

 

[Rainbow Man]

Presumably though in networked render system like this, you'd be using the processing power to add more detail rather than render faster.

Yes that does make a lot of sense. Thanks!

Would the 256MB (512 counting video memory) be a big hindrance for professional renderfarming you think?

Surely any single PS3 wouldn't have to keep that much texture data etc in memory if it's just rendering a small section of the screen.

All sytstems could have all textures and such on their harddrives and load only the pieces that will be visible in their respective section scould they not?


Peace.

 

[Shifty Geezer]

Yes that does make a lot of sense. Thanks!

Would the 256MB (512 counting video memory) be a big hindrance for professional renderfarming you think?

Absolutely. You'd be nuts to try and build a professional network farm on PS3s though! You'd buy Cell blades and not have hundreds of pointless BRD drives and RSXs cluteruing up the office.

All sytstems could have all textures and such on their harddrives and load only the pieces that will be visible in their respective section scould they not?

That's certainly what you'd do to elliminate a lot of network traffic, rather than stream the textures from the server as needed.

The whole RT algorithm doesn't make much sense though. You wouldn't build a render farm to raytrace only. You'd want to run different algorithms, and the pro's stick with RenderMan. Dunno how well that would port to Cell.

 

[Megadrive1988]

seriously I would love to see what a bunch of PlayStation3s can pull off when a.) not having to RT, b.) when both RSX and CELL are used together, not just software-based rendering on CELL.

 

[sunscar]

I said DEcompression. As in a single SPE in the destination machine. The one that produces th efinal monitor output image.

Peace.

There's no reason they should be limmited to only one SPE, it's likely a case-specific choice, but sorry for the misread there.

 

[sunscar]

seriously I would love to see what a bunch of PlayStation3s can pull off when a.) not having to RT, b.) when both RSX and CELL are used together, not just software-based rendering on CELL.

Absolutely didn't mean to post a double, but I missed this post.

Terra Soft (The company responsible for PS3 Linux) has stated publicly that they're pressing SONY to allow access to the RSX (And consequently, access to its 256MB chunk of GDDR3), so it's atleast remotely possible we'll see this in a few years.

Regardless of if or not it becomes a reality, I'd still like to get five or ten of these together to run a few simulations on, just to get an idea what kind of power we'll have access to in a single box in a few years.