Use ps3 as a workstation for 3d?

[Laa-Yosh]

Alias had their beta cell based workstations last year.They allready worked on intensive and parallelized realtime cloth simulations.

Maybe they've decided to rewrite their cloth stuff, now that Syflex has almost totally overtook the market?

BTW, where did you learn about this?

 

[Titanio]

Theoretical question here: if linking together PS3s to pool processing power works out of the box, as suggested, would the amount of RAM be additive? Or would it be more akin to SLI cards from that perspective, where the effective amount of RAM isn't doubled with a second card?

 

[Bobbler]

Theoretical question here: if linking together PS3s to pool processing power works out of the box, as suggested, would the amount of RAM be additive? Or would it be more akin to SLI cards from that perspective, where the effective amount of RAM isn't doubled with a second card?

It should be handled like a standard NUMA situation (adding more PS3s should act like adding more total memory as the Cells should be able to dip into everyone elses pools with the way Cell was supposedly built) -- but I imagine it wouldn't be all that fantastically fast, seeing as it has to go over a network connection (which at best is around 100mb/s on a gigabit network).

 

[Shifty Geezer]

Maybe they've decided to rewrite their cloth stuff, now that Syflex has almost totally overtook the market?

BTW, where did you learn about this?

Earlier in the year we all had a nose at an Alias paper on Cell's cloth simulator performance. If the OS included on PS3 HDD (assuming it is) is compatible with the Cell workstations Sony want to create that presumably Alias are writing for then I'd guess Maya would be usable. I'd love to see what the performance is like, and how algorithms improve over time.

 

[Laa-Yosh]

We could certainly use a faster machine for cloth work...

 

[Bobbler]

Earlier in the year we all had a nose at an Alias paper on Cell's cloth simulator performance. If the OS included on PS3 HDD (assuming it is) is compatible with the Cell workstations Sony want to create that presumably Alias are writing for then I'd guess Maya would be usable. I'd love to see what the performance is like, and how algorithms improve over time.

Didn't that paper show with their initial attempt they were able to put together something that outperformed a P4 by something like 5x? (I can't remember exactly, maybe someone can dig up the article)

 

[Titanio]

Didn't that paper show with their initial attempt they were able to put together something that outperformed a P4 by something like 5x? (I can't remember exactly, maybe someone can dig up the article)

8 SPUs at 2.4Ghz were ~5x faster than a P4 @ 3.6Ghz.

It was a port, not built from the ground up for Cell, too. They took the perhaps easier option of running 8 different cloth simulations across the 8 SPUs versus 1 larger simulation across all 8.

 

[_phil_]

http://www.research.ibm.com/cell/whitepapers/alias_cloth.pdf

btw,it's new solver algorithm.

 

[Shifty Geezer]

8 SPUs at 2.4Ghz were ~5x faster than a P4 @ 3.6Ghz.

It was a port, not built from the ground up for Cell, too. They took the perhaps easier option of running 8 different cloth simulations across the 8 SPUs versus 1 larger simulation across all 8.

I contacted Alias about this and they said it was a prototype cloth solver algorithm written for Cell using single precision. Each SPU was about half as fast as a P4, except the first which was 1x the speed, with the PPE contributing next to nothing, which was odd. Performance per square mm of silicon was 5x higher than a P4, but overall I would have hoped for better, given that this is an area Cell is supposed to accel at. But then again I don't know what level of accuracy etc. they were aiming for.

_phil_'s link covers the paper nicely...

This technology demonstration shows a prototype of a next generation cloth solver algorithm. It is a proof-of-concept that demonstrates the potential for interactivity and realistic display of cloth behaviors. The demonstration shows independent simulations being added, one by one, with no degradation in performance.

Not really an attempt at a best case solver, rather than an experiment, it seems.​

 

[Titanio]

I contacted Alias about this and they said it was a prototype cloth solver algorithm written for Cell using single precision. Each SPU was about half as fast as a P4, except the first which was 1x the speed, with the PPE contributing next to nothing, which was odd. Performance per square mm of silicon was 5x higher than a P4, but overall I would have hoped for better, given that this is an area Cell is supposed to accel at. But then again I don't know what level of accuracy etc. they were aiming for.

The info you got conflicts with the paper a little. The paper explicitly talks about porting to Cell vs writing something just for it, for one. The first SPU is also more than 1x as fast as the P4 (well, just a little more). There isn't a linear speedup, which I think is a better way of saying that the subsequent SPUs aren't as fast as the P4, individually. That said, if you linearly scaled the performance to match the P4's clockspeed, you'd be a little closer to that linear speedup (but I know it wouldn't necessarily scale like that..but it may).

 

[Laa-Yosh]

They took the perhaps easier option of running 8 different cloth simulations across the 8 SPUs versus 1 larger simulation across all 8.

Cloth might be hard to paralelize, as each vertex depends heavily on the surrounding vertices - they're all continously connected with springs.

 

[Laa-Yosh]

Regarding calculation times, our cloth stuff usually takes a few hours to simulate for a few seconds of animation at 24fps, on a 3.2 Ghz Xeon. Only heavily complicated scenes can take a whole night... However, complexity is dependent on totally different factors, like the stiffness of the cloth, speed of motion, use of self-collision etc.
Oh, and we're using the Syflex plugin and not the builtin Maya solver.

 

[aeriic]

That said, if you linearly scaled the performance to match the P4's clockspeed...

Why would you do that? It's not very common for parallel algorithms to have linear O speedup. The only way you can do that is if you assume the parallel overhead is negligible, which I highly doubt...

 

[frio]

Cloth might be hard to paralelize, as each vertex depends heavily on the surrounding vertices - they're all continously connected with springs.

Laa-Yosh, you are correct, the algorithm is non-trivial to parallelize, although the most intense part is computing collisions rather than springs...

Cheers,

Frio

 

[frio]

The info you got conflicts with the paper a little. The paper explicitly talks about porting to Cell vs writing something just for it, for one. The first SPU is also more than 1x as fast as the P4 (well, just a little more). There isn't a linear speedup, which I think is a better way of saying that the subsequent SPUs aren't as fast as the P4, individually. That said, if you linearly scaled the performance to match the P4's clockspeed, you'd be a little closer to that linear speedup (but I know it wouldn't necessarily scale like that..but it may).

Titanio, another reason for the non linear speedup is that the PPE participates in the computation for each simulation that is running on the SPUs.
Anyway increasing the clock speed would proportionally increase the overall computing speed, but it would still not generate a performance directly proportional to the number of SPUs.

Cheers,

Frio