[ODE] Your thoughts on "Physics Accellerators"

Nate W coding at natew.com
Mon Jul 1 22:47:02 2002

On Mon, 1 Jul 2002, Russ Smith wrote:

> in other words - go out and buy a better CPU! (or install some nice
> fast DSP hardware).

So maybe "DPU card" is just another way of saying "DSP card" but with some
extra stuff...

* lots of fast onboard RAM, as you said
* A standard API for dynamics programming - call it OpenDL :-)
* DMA (as you said) to move data from CPU-RAM to DPU-RAM and back again
  (or from CPU to DPU to GPU, but that would take some creativity...)
* An SDK to allow dynamics libraries to be ported to the DPU's DSP with as
little overhead as possible.

Not to mention lots of marketing to get dynamics people and application
people to try porting their libraries and titles to OpenDL.  If you could
get multiple DPU/DSP hardware vendors to play leapfrog like AMD/Intel and
nVidia/ATI/etc, these cards might get fast, fast.

The first video game to hit the shelves using a DPU could have (I figure)
a physical environment with at least twice the dynamic behavior as
anything else, plus a lot of extra CPU time for more interesting AI or
whatever.  The same crowd that buys $400 video cards might give it serious

There's kind of three-way chicken-and-egg problem though - users won't buy
DPUs until games support them, games won't be written for DPUs until DPUs
are available, and DPUs won't be available until users demand them.  I
ain't holding my breath for this situation to resolve itself.

For the last couple years I had been thinking that CPUs have pretty much
gotten 'fast enough.' Messing with ODE changed my mind about that - I
don't mean that as critcism of ODE of course, it's just that realistic
physics requires so much computation that hot computers are no longer fast
enough.  DPUs could solve that problem pretty quickly, and there might be
a business case in the gaming market to make them worthwhile.  


Nate Waddoups
Redmond WA USA