SMART: a neurocomputer using sparse matrices

SMART (Sparse Matrix Adaptive Recursive Transforms) is a new architecture supporting sparse matrix transforms. Though it is primarily designed for neural network simulations, the approach used is general enough to encompass other domains (e.g. signal processing). SMART efficiently implements algorithms that use a vectorial formalism. A simple sparse matrix compaction procedure is given which parallelizes computations thanks to dedicated data structures. Using this architecture, we design a vector processor that interfaces with a host machine which is thus relieved from intensive computational tasks.<>