Abstract
The continuous miniaturization of CMOS feature sizes into the nanometer regime has increasingly caused problems due to noise vulnerability, process variations, and energy consumption. Noise vulnerability and process variations have been recognized to cause statistical or probabilistic device behaviors. In this paper, by capitalizing on what most people termed as nuisance, we demonstrate how noise can be put to good use in CMOS devices. We present the first hardware implementation of a Bayesian medical diagnostic system for real-world patient monitoring using probabilistic CMOS (PCMOS). We explore ways to adapt Bayesian structure under realistic hardware constraints without compromising prediction accuracy. As compared to conventional CMOS, PCMOS implementation of the network offers noteworthy merits: ultra low-power consumption, high-speed performance and cost-effectiveness.