A GENERAL-PURPOSE SIGNAL PROCESSOR ARCHITECTURE FOR NEUROCOMPUTING AND PREPROCESSING APPLICATIONS

A general-purpose Neural Signal Processor MA16 is presented the architecure of which is guided by an analysis of today's neural algorithms. The MA16 executes the elementary algorithmic strings which are compute-bound and shared by all neural nets; operations which are not time consuming are let to hardware off-the-shelf. Digital design is chosen because of flexibility and computation accuracy. The throughput is 800 million connections per second at a clock frequency of 50 MHz (I connection = 16 bit). This performance is valid for arbitrary networks provided they consist of more than 16 neurons each comprising more than 16 synapses. Depending on the needs of the application under consideration a linear or 2-dimensional array of VLSI chips can be constructed in order to provide sufficient processing power and weight memory.