The Scientists' Opinions on Gaming Physics

Erleben actually prefers the GPU solution, because GPUs are quite flexible to program now. "With the GPU, one has full control over the GPU hardware. The GPU offers two other interesting aspects: it is a streaming architecture, and it is inherently parallel." He also, however, talks about some of the issues connected to working with graphics accelerators, like the nearly impossible task of debugging code and vendor-specific issues. We guess that part of the reason he prefers working on the GPU is that these devices have been around a lot longer; in a few years we may hear the same thing about the PPU, too...

According to Lacoursière, there is one main problem with the iterative (GPU, PPU) solutions that are used instead of the direct, more accurate ones. Since they all work by making rough approximations, they don't take into account any long-term effects at all, which would require too much computing power. This, says Lacoursière, is the reason why both the PPU and GPU fail in being a sufficient platform. This job should instead be handled by the master of solving direct matrix operations: the CPU. The raw processing power of dual multi-core CPUs with large caches should mean that there is no problem in improving performance a great deal. Still, as Erleben points out, to really make good use of this multithreading architecture, you need to redesign all the algorithms involved, which will take some time to do.
http://tomshardware.co.uk/2006/08/03...ng_physics_uk/