Analog in-memory computing, sometimes shortened to AIMC, is a way of building AI hardware that performs the multiplications at the heart of a neural network directly inside the memory cells that store the weights, using physics rather than digital arithmetic. In a conventional computer, weights live in memory and have to be fetched into a processor to be used; with large models this back-and-forth - the von Neumann bottleneck - consumes most of the energy. In-memory computing collapses the two together.
The trick relies on basic circuit laws. If each weight is stored as the conductance of a memory cell arranged in a grid, applying input voltages along the rows produces currents that add up along the columns by Kirchhoff’s law, and the resulting current is the weighted sum the network needs. A whole matrix-vector multiplication - the dominant operation in deep learning - happens in one analog step, in place, without moving the weights anywhere. Phase-change memory and memristor devices are the usual cells used to hold the analog weight values.
IBM has been the most public champion. In 2023 it described an analog AI chip built from phase-change memory, with 64 in-memory compute cores each holding a 256-by-256 crossbar of synaptic units, and reported running speech-recognition and image tasks at accuracy comparable to digital chips while being many times more energy efficient. A related IBM result published in Nature reported over ten times the energy efficiency on a speech benchmark.
Why business readers should care: the cost of running AI is increasingly an electricity bill, and analog in-memory computing is one of the most concrete bets on a more efficient substrate for inference. The hard part is accuracy: analog cells drift and vary, so getting digital-level precision at scale is the engineering battle that decides whether this leaves the lab.