An FPGA-based Probability-aware Fault Simulator

A recent approach to deal with the challenges that come along with the shrinking feature size of CMOS circuits is probabilistic computing. Those challenges, such as noise or process variations, result in a certain probabilistic behavior of the circuit and its gates. Probabilistic Computing, also referred to as pCMOS, does not try to avoid the occurrence of errors, but tries to determine the probability of errors at the output of the circuit, and to limit it to a value that the specific application can tolerate. Past research has shown that probabilistic computing has potential to drastically reduce the power consumption of circuits by scaling the supply voltage of gates to a value where they become non-deterministic, while tolerating a certain amount of probabilistic behavior at the output. Therefore, one main task in the design of pCMOS circuits is to determine the error probabilities at the output of the circuit, given a combination of error probabilities at the gates. In earlier work, pCMOS circuits have been characterized by memory-consuming and complex analytical calculations or by time-consuming software-based simulations. Hardware-accelerated emulators exist in large numbers, but miss the support of injecting errors with specified probabilities into as many circuit elements the user specifies at the same time. In this paper, we propose an FPGA-based fault simulator that allows for fast error probability classification, injection of errors at gate-and RT-level, and that is furthermore independent on the target architecture. Moreover, we demonstrate the usefulness of such a simulator by characterizing the probabilistic behavior of two benchmark circuits and reveal their energy-saving capability.