Cuda/OpenCL

Hiya, quick question. As and when you tackle OpenCL or Cuda, how much of a ballpark performance increase do you think you could get over CPU cores? I'm thinking it might be much better bang for buck, to have a 2nd GPU card dedicated to the physics, rather than trying to max out expensive CPU cores. An amazing number of cores on mid range GPU's these days, which are generally highly suited to physics?

 

You can try to look up some papers about parallel computing of mass-spring systems like this. It's quite a challenge to transform your data to an array of independent bits. Take a beam, it connects a pair of nodes. But these nodes are connected to others with more beams. So you have several processors (each dealing with a single beam) that need access to the same memory address simultaneously. Not good. You can preprocess your data and put it back together at the end, but the overhead can easily defeat advantages of GPU computing. Also you could forget about additional features that need access to the node/beam data every physics frame, because you just can't copy everything from RAM to VRAM and back 2000 times per second. Etc...
Moreover graphics like this needs some power too.

 

It was an extract from a IEEE paper that suggested you could get a decent performance boost.

"specifically a highly realistic mass-spring model, and a GPU based technique, and analysis, that provides a nearly 80x speedup over serial execution and 20x speedup over CPU based parallel execution."

I'll maybe just buy the paper.

 

The "specifically a highly realistic mass-spring model..." part suggest that the GPU would be busy for a long enough time with each data element that it will compensate for added overhead, like when you send the GPU some command which causes a stall.