Some more beef on matrix (or problem) sizes and performance:
Citing from
http://vtb.engr.sc.edu/vtbwebsite/downlo...a%20Methods.pdf
"We have used Graphics Processing Units (GPUs) to acceler- ate the solution of the types of equations typically encoun- tered in dynamic system simulators. Compared to commer- cial matrix solvers that run on a CPU, we realized speedups ranging from 5 (for system size ≈ 700) to 460 (for system size ≈ 5800)."
I estimate MAME netlists to be well below a system size of 100. Thus overhead would kill performance. We are calculating realtime sound, i.e. at a rate of 48000 Hz. Thus you need to pass a couple of matrices 48000 times a second to the GPU. I assume, that the transfer CPU-GPU is designed to cope with Screen Refresh frequency (order 50 Hz).
But if interest (or need) arises, the netlist model abstraction can easily be extended to use a GPU solver instead of the Gauss-Seidel and Gaussian Elimination solvers currently used.
|