[ODE] hw

Steve Dekorte steve at dekorte.com
Sun May 11 15:08:01 2003


On Sunday, May 11, 2003, at 10:00 AM, Adam Moravanszky [NovodeX] wrote:
> I wrote an article for the upcoming book ShaderX 2 about some GPU based
> linear algebra solvers, including an LCP solver (which is what you  
> need to
> do constrained rigid body simulation).  These run on any graphics card  
> with
> decent DirectX 9 support, and I clocked them to be competitive with or
> faster than SSE2 optimized CPU code.

Hi Adam,

Thanks for the response. That sounds interesting.

> There is also a paper coming in
> Siggraph 2003 about sparse solvers, which are even more interesting.   
> So
> clearly, it is not neccesary to create yet another piece of specialized
> hardware for the PC -- with GPUs you already have/will have powerful,
> general purpose floating point vector processors.

I think it depends on the cost/performance ratio(and silicon is cheap).  
For example, most folks find it worth adding 10% to the cost of their  
PC to have 1000x the graphics performance (the primary use of which is  
games). I suspect the same will be true of dynamics performance. For  
example, this paper:

http://web.informatik.uni-bonn.de/II/ag-klein/people/zach/papers/ 
wscg.html

describes a first generation collision detection hardware that appears  
to be 10-200x faster than CPU based approaches. I suspect the  
in-quantity cost will be much less that of graphics hw, which is now in  
the $50/PC range for (I'm guessing) >1000x CPU performance.

Cheers,
Steve
"Statically typed languages are like American sports cars.
They go fast, but only in a straight line"