An Ultra-Compact Calculation Unit with Temporal-Spatial Re-configurability

We develop an ultra-compact calculation unit with temporal-spatial re-configurability. On the basis of a novel bisection neural network (BNN) topology, a processing element array is flexibly partitioned into multiple calculation units on the hardware. Re-configurability in spatial (function) is achieved by adjusting the shape and location of each calculation unit. Meanwhile, we integrate stochastic computing (SC) logic on the proposed calculation unit to take advantages of the BNN topology. Re-configurability in temporal (accuracy) is achieved by adjusting the length of SC bit-stream. Experimental results show that our calculation unit outperforms some state-of-the-art approximate calculation units in terms of energy efficiency.