Running a neural network requires moving many numbers between memory and processors. In-memory computing performs some operations where those numbers are stored. It describes an architectural approach rather than a particular memory technology, and implementations may be digital or analog.
In an analog array, cell conductance can represent weights. Applied voltages represent inputs, while currents combine according to circuit laws. The arrangement performs an operation resembling matrix-vector multiplication, a major component of neural-network workloads. IBM research uses phase-change memory among its approaches: a material’s physical state changes its electrical properties.
Reduced data movement could benefit AI accelerators and energy-limited devices. Input, conversion and unsupported operations still require surrounding circuitry. Noise, cell variation and numerical precision affect results. Evaluating a complete system on a defined task is therefore essential. Memristive devices can provide building blocks, but the terms are distinct: a memory device is a component, while in-memory computing describes its computational use.