Eliminating a hidden error source in stochastic circuits

Stochastic computing (SC) computes with probabilities using random bit-streams and standard logic circuits. Its advantages are ultra-low area and power, coupled with high error tolerance. However, due to its randomness features, SC's accuracy is often low and hard to control, thus severely limiting its practical applications. Random fluctuation errors (RFEs) in SC data are a major factor affecting accuracy, and are usually addressed by increasing the bit-stream length N. However, increasing N can result in excessive computation time and energy consumption, counteracting the main advantages of SC. In this work, we first observe that many SC designs heavily rely on constant inputs, which contribute significantly to RFEs. We then investigate the role of constant inputs in SC, and propose a systematic algorithm CEASE to eliminate them by introducing memory into the target circuits. We provide analytical and experimental results which demonstrate that CEASE is optimal in terms of minimizing RFEs.