PPU = good for vertex processing?

[Brimstone]

Some ballpark performance numbers that Ageia has thrown out.

In rigid-body dynamics, for example, the chip must solve the Newtonian physics problems common to most games. Today's games track and solve the interactions between perhaps 10 to 200 rigid bodies using software simulations, Hegde said. The PhysX chip will support up to 32,000 rigid bodies, all interacting with each other, he said.

While graphics chips are measured by the number of pixels they draw per second or the number of polygons they can trace, Hegde said that the PhysX chip will be compared against a suite of measurements, generally in relation to a purely software simulation. Rigid body dynamics, for example, will be marked by the number of collisions per second that can be traced, while the chip's performance calculating fluids will track the number of particles used to model the fluid, and how each particle interacts with its neighbors. Modeling clothing also requires the chip to track the number of particles and their collisions with neighboring particles and other objects, similar to the way the chip will model hair.

Hegde declined to describe the chip's performance in detail. In addition to the 32,000 rigid bodies the chip will model, the PhysX chip will be able to track 40,000 to 50,000 individual particles as part of a fluid. While impressive, that's nowhere near the 10 million particles fluids simulated in feature films like The Lord of the Rings use, he said.

A Novodex-enabled game will sense the presence of the PhysX chip and use the technology, Hegde said, and offload the physics burden from the other two chips. In the case of awar game, for example, a rocket launcher may blow a building into 10,000 pieces, some of which might interact with other building, the player, other players, or a myriad of other scenarios, he said.

http://www.extremetech.com/article2/0,1558,1773957,00.asp?kc=ETRSS02129TX1K0000532

 

[scificube]

What was meant exactly by off loading the burden from the other two chips?

Perhaps this is a clue as to how the PPU functions.

The CPU is obviously one of them...would the other chip be the GPU and it's vertex processing or at least some of it....

What else would make sense? I don't know how 3D positional sound etc works on sound cards but isn't it just make x sound when the CPU tells it to...maybe some positional data supplied by the CPU, but the sound card does the work for the effects right?

Somebody in the know better than...could you please help me out?

edit:

One thing that has come back to me is that Hedge could be referrig to how some were attempting to or make the GPU a viable candidate to perform physics interactions such as fluid dynamics.

In this sense a PPU would remove such a load from the GPU or in truth eliminate the need to purse trying to put such a load on the GPU in the first place. Maybe this could be what Hedge meant.

 

[IgnorancePersonified]

Sounds like good arguments to intergrate a PPU style unit onto a CPU.

 

[scificube]

Sounds like good arguments to intergrate a PPU style unit onto a CPU.

How so?

Then...you wouldn't be offloading the load from the CPU. You'd be improving it if you could justify adding so many transistors to a general purpose chip for such specialized tasks. (This would include making the cache larger as well)

I really don't think that would be such a good idea. For those who don't game or work with 3D graphics I suppose the PPU unit would be a whole lotta wasted space that still made them pay more for such a CPU.

I think it works better as a standalone unit in that you'll buy it if you need it and thus you'll get you're money's worth. (Unless you just bought it to be cool or something...) If you don't need it, you don't buy it and you don't waste money.

 

[nutball]

Hehe, lets return to the x86 + x87 co-pro days

 

[Frank]

Those particles for the volumes, would they look like a 3D mesh (ie. not only the outside)? If that is the case, they might generate the new 2D mesh by simply removing the inner particles. I'm assuming they don't just return new positions for anchor points and let the CPU calculate the new meshes, as that would be really inefficient IMO.

 

[scificube]

Hehe, lets return to the x86 + x87 co-pro days

Dude, I love your sig!