Game Development Reference
number of solver iterations is specified as 10. All rigid bodies are always moving,
and the maximum number of constraints is about 20,000. We can easily imagine
how difficult this scene would be for both broad phase and constraint solver. See
Figure 3.17 for the result. The vertical bars show the time of computation in a
simulation frame. The optimized version with five SPUs is about 12 times faster
than the original PPU version. It shows that we can achieve a large performance
improvement with the effective combination of parallel computation and double
buffering on multiple SPUs.
Note that computation on a single SPU is three and one-half times faster than
on a PPU; that is, it shows that even a single SPU without parallelism is faster
than a PPU.
Figure 3.17. The result of the performance.
[Catto 05] Erin Catto. “Iterative Dynamics with Temporal Coherence.” Technical
Report, Crystal Dynamics, Menlo Park, CA, 2005.
[Ericson 05] Christer Ericson. Real-Time Collision Detection . San Francisco:
Morgan Kaufmann, 2005.