Hi, my application needs to compute singular values of several hundred 30x30 matrices concurrently, in periodic batches. I am prioritizing speed/throughput over accuracy - I was wondering which one I should explore using.
Does anyone have any resources/experience for comparing the three? Am I missing any options?
this may be of interest. There may be no particular application to the batched functions, however.
cuSolver is what you want to look into. It solves up to 32x32 in batches, supporting different precisions and also complex numbers. Definitely use the batched APIs like gesvdjBatched, syevjBatched
As far as svd algorithms are concerned, Jacobi (Gesvdj) works best on small matrices (not sure if 30x30 still qualifies as small).