Compute-in-memory / analog in-memory computing

Compute-in-memory (CIM) performs the dominant operation of neural decoding — the multiply-accumulate of weights and inputs — inside the memory array that stores the weights, rather than fetching weights to a separate arithmetic unit. In the analog form, weights are programmed as device conductances in a crossbar; applying input voltages produces, by Ohm's and Kirchhoff's laws, a summed current that is the dot product in a single step. This collapses the data-movement cost that dominates digital inference and can raise energy efficiency by orders of magnitude for matrix-heavy workloads.

The costs are analog: finite precision, device-to-device and cycle-to-cycle variability, conductance drift, and the energy and area of the edge data converters (ADCs/DACs), which often dominate a real design. For BCI these constraints must be met at implant scale and reliability, so most demonstrations remain benchtop chips evaluated on neural benchmarks rather than chronic devices.