A neural network integrated circuit supporting programmable exponent and mantissa

A neural network architecture is presented utilizing small (8-b) word sizes for data and weights while maintaining flexibility in the dynamic range achieved by using adjustable exponent and mantissa field. Design issues regarding tradeoff between small and large precision number systems are reviewed. Simulation results on network performance are included. In small precision systems (8-b), a fixed dynamic may limit the system capability. For certain applications it is desirable to have a flexible dynamic range implementable with a variable size of exponent and mantissa