A bit of info on Cell's physic's abilities.
- Thread starter scatteh316
- Start date
The rendering of the mesh visuals has no impact on the experiment which is a cloth simulation solver. Whether they connect to a Mac, PC, or Silicon Graphics workstation, it's still only the Cell that's processing the simulation.Lysander said:
The Blade Server prototype presumably has no GPU, it being a server, so the data from the simulation will need to be passed onto a rendering platform of some sort to visualise the results.
Ummmm, not sure what you mean. The processor works out simulations and can create models and textures and stuff procedural, and then that gets rendered on the GPU. In that respect it's kinda similar. But really this is more about the offline rendering industry. A CGI artist using Maya will create their models on their desktop workstation along with cloth content. They'll then want to preview the motions of the cloth to see how then animation looks. That'll take maybe 5 minutes on a 3.62 GHz P4, whereas if they were working on a Cell workstation, or had a Cell server that they could get to do the work, the preview would be done in 1 minute. Or maybe the difference is a 10 frame a second preview on the P4 and a 50 fps preview with a Cell instead.
The presence of the Mac was just to show the results of the Cell's calculations. In a Cell workstation the Mac wouldn't be needed as the Cell would do the calculations and it's internal GPU would show the results.
The presence of the Mac was just to show the results of the Cell's calculations. In a Cell workstation the Mac wouldn't be needed as the Cell would do the calculations and it's internal GPU would show the results.
ShootMyMonkey
Veteran
All the same, 40,000 simulation points is not that lightweight, especially when you're actually dealing with self-intersection. Hell, if it's achieving the hypothetical 50 fps you pose on that many points, that's actually approaching the performance the AGEIA's PPU. I'd like to know more about how it actually performed. Collision against other objects is the easy part, self-collision is so-so, but getting it all to look good means doing it all on a hell of a lot of points, and that's the killer.
In a real game situation, you really wouldn't deal with that many points on just 8 pieces of cloth. But if the performance is decent with that kind of granularity on not-so-well optimized code, I think that bodes well.
ShootMyMonkey said:
Thats the general conception i'm getting from this performance. A new CPU comes out, trumps a top of the line P4 right of the bat and the code may or may not be optimized for it.
ihamoitc2005
Veteran
Can it be?
A few surprising items that stood out in Alias article:
1. SPE runs generic code faster than PPE. (1 SPE or all SPEs together?)
2. VMX has very poor performance compared with P4 or 1 SPE in this simulation.
3. 1 SPE more powerful than P4 in simulation but each extra SPE adds maybe 50-60% of P4 capability in more or less linear amount.
I read thread so I understand statement about Alias = not so fast coding etc. What is curious is VMX performance. I expected VMX = SPE performance. Xenon rated at 115Gflops, or 1.5 x SPE per core. This is also why CELL rated at 218Gflops, because PPE is similar to Xenon core and rated at 1.5 x SPE.
A few surprising items that stood out in Alias article:
1. SPE runs generic code faster than PPE. (1 SPE or all SPEs together?)
2. VMX has very poor performance compared with P4 or 1 SPE in this simulation.
3. 1 SPE more powerful than P4 in simulation but each extra SPE adds maybe 50-60% of P4 capability in more or less linear amount.
I read thread so I understand statement about Alias = not so fast coding etc. What is curious is VMX performance. I expected VMX = SPE performance. Xenon rated at 115Gflops, or 1.5 x SPE per core. This is also why CELL rated at 218Gflops, because PPE is similar to Xenon core and rated at 1.5 x SPE.