Approximate Computing in Memristive Processing-in-Array (PIA)

Memristor-based in-memory computing has been the topic of much research and development in recent years. Among them, Processing in Array (PIA) boasts of high energy and delay efficiency, since in this type of memory-centric computing, the data never leaves the array for being processed. PIA literature is divided into two major areas; i. analog computing (mainly for matrix multiplication) and ii. digital logic for general-purpose computing. The latter, which is conducted using commonly known stateful logics, despite its versatility and high parallelization ability, requires many steps for each (sub)function. In this work, we present approximate PIA for memristive stateful logic, and how it can significantly reduce the number of steps required for each function and consequently decrease the energy consumption, delay, and potentially the number of required memristors for computation. Approximate computing introduces errors into computations, which, as we show, can be tolerated in error-resilient applications such as image processing and machine learning.