Hardware Efficiency Stochastic Computing based on Hybrid Spatial Coding

As the era of silicon-based microchips advances by Moores Law approach to physical limits, new computational paradigms are proposed for future systems, i.e., stochastic computation. However, the current stochastic computing faces the challenge of high latency and low accuracy. In this work, we propose spatial coding based on the hybrid stochastic computation (SHSC) method, which is a stochastic-binary hybrid domain computation. Instead of sequential bits computing, the proposed SHSC expands stochastic bits in the spatial dimension. To balance the accuracy and complexity, the multiplication is divided into high and low precision parts, where the high precision parts are performed in the binary domain, and low precision parts are performed in the stochastic domain. A low-cost error compensation circuit is proposed to further improve the computation accuracy. According to the implementation outcomes, the proposed method exhibits a 28% hardware efficiency improvement with the same inference accuracy as traditional neural network applications.