[ODE] Your thoughts on "Physics Accellerators"
Russ Smith
russ at q12.org
Mon Jul 1 21:59:02 2002
> I think there's a lot of potential in that idea. I could see a DPU
> (dynamics processing unit) alongside the CPU and GPU in a high-end
> gamer's machine, and probably other applications as well. My guess is
> that the"DPU" idea would work best if the processor was designed to
> work well with a variety of dynamics software libraries though, not
> just ODE or just Dynamechs or just any particular library.
>
> If a "DPU" did for dynamics what GPUs have done for graphics, we could
> be simulating very complex worlds in real time, with hundreds or even
> thousands of bodies and joints. That would be too cool.
the problem with this is that dynamics tends to be more of a 'software'
problem than a 'hardware' problem. what i mean by this is that dynamics
algorithms are much more subject to changes based on the requirements of
a particular simulation. the computational expense is mostly in the
matrix manipulation (small and large), but there are a wide variety of
different matrix algorithms that might be used (e.g. different
factorizations, incremental factorizations, sparse matrices with known
and unknown patters). the common operation in graphics is 'draw this
pixel', and the hardware speedup comes because this is (relatively) easy
to express in hardware and can be massively pipelined. the only common
element in matrix operations is the multiply-add. so, rather than a
special purpose dynamics chip, what we actually need is a chip that can
do more/faster multiply-adds, and more importantly a chip that can feed
the data to the MADD units as fast as possible. in other words, the wish
list is:
* SIMD MADD units (as multi-way as we can get)
* fast scratch RAM
* low latency memory heirarcies and/or fast programmable DMA
to get the data as fast as possible.
these are the requirements for *any* intensive numerical task, not just
dynamics. in other words - go out and buy a better CPU! (or install some
nice fast DSP hardware).
russ.
--
Russell Smith
http://www.q12.org