Neurotransputers: new hardware component for parallel neurocomputers
K.V. Chernukhin, M.A. Samarin, E.A. Gryasin · 2002
Specific solutions in transputers' architecture make their application in neurocomputing remarkably effective in comparison to traditional equipment, because they offer hardware parallel computing support and serial data transfer channels. But simulation of the neural networks with many connections may require more communicational resources then usually present in transputers. Furthermore, transputers' orientation to the wide spectrum of tasks makes their mathematical operations to be redundant for neurocomputing. Hence, the development of new hardware component for software and hardware simulation of neural networks is reasonable. The authors develop neurotransputer hardware components for parallel digital neurocomputers and necessary software. The neurotransputer developed consists of the following functional units: arithmetic-logical unit, which supports mathematical operations, required for algorithmic representation of the neuron's models; specialized coprocessor for fast output functions calculation; memory units with upgrade features; serial data transmission links; base neural unit (BNU), for preliminary data processing and extension of communicational functions of serial links.