[ODE] hw

[Russ Smith]

> Note that this presents sparse methods for systems of linear equations, not
> LCPs.  My paper also covers PSD LCPs, but for dense matrices, rather than
> sparse.  I haven't read the above paper in detail yet, but I think that its
> possible to extend it to sparse LCPs.

hi adam,

thanks for the reference. interesting paper. on the subject of hardware
accelarated physics my own feelings tend towards "give me a faster
general purpose processor". of course specialized hardware (GPU, PS2 SIMD,
whatever) is a big win when the algorithms you're running are reasonably
simple and stable. but with any such hardware there's going to be a big
investment in optimization that not only ties you to the hardware platform
but also to the algorithms you're using. a hardware accelerated dense
matrix method is no good for an iterative LCP method or a reduced
coordinate method (and vice versa). i believe that truly interesting
simulations are going to be multi-domain and will require diverse
numerical methods to work - which requires a very flexible (and therefore
somewhat general purpose) computational engine. but obviously hardware
accelerated specialized algorithms are always going to fill a
valuable niche.

one other thing to note is that mass market general purpose chips
(pentium, athlon, power, GPUs etc) are usually ahead of the game
process-wise and can therefore be way faster than more specialized
hardware. for example, with BLAS you can get something like 80-90%
efficiency on these processors with dense matrix operations - this can be
way more flops than you'd ever get out of your FPGA. just a thought.

russ.

--
Russ Smith
http://www.q12.org/

On Mon, 12 May 2003, Adam Moravanszky [Novodex] wrote:

>
> --  Adam Moravanszky
> /*================*\
> |Chief Software Architect
> |NovodeX AG
> |physics middleware
> |www.novodex.com
> \*================*/
>
>
> _______________________________________________
> ODE mailing list
> ODE@q12.org
> http://q12.org/mailman/listinfo/ode
>