Compute in Memory

last updated 2026-08-31

Physics / mechanism

Compute in memory places arithmetic where data already resides, removing the round trip between a separate memory array and a logic unit. The canonical primitive is a resistive crossbar: programmable conductances hold matrix weights, applied voltages act as the input vector, and the summed currents on each column perform a multiply-accumulate in the analogue domain. Non-volatility matters because weights must persist without refresh or standby power, so the array can be read repeatedly at low energy per operation.

The photonic variant of this idea is the subject of the available source material. Photonic integrated circuits offer large bandwidth, low latency and inherent parallelism for communication, sensing and information processing, but lack efficient, scalable, non-volatile memory elements on chip. Opto-electronic resistive memories are proposed as the missing element: a device whose resistance state, and hence optical response, can be set electrically or optically and retained without power, allowing weights to be stored in the same structure that modulates the light carrying the data.

Key parameters for any such device family are the number of distinguishable conductance or transmission states, retention and endurance, switching energy, the insertion loss added to the photonic path, and whether programming is compatible with foundry back-end processing. The source frames the problem as one of efficiency, scalability and non-volatility simultaneously, which implies that partial solutions on any one axis have not been sufficient.

Competitive landscape

The sources support only a narrow comparison: electronic resistive memory arrays versus opto-electronic resistive memories embedded in photonic integrated circuits. The photonic route is motivated by bandwidth, latency and parallelism advantages of the optical domain, and is bottlenecked by the absence of a suitable non-volatile memory element rather than by the arithmetic itself. No comparative performance data across device stacks is present in the supplied material.

For funding context, semiconductor startup capital in the quarter was concentrated in AI hardware, with edge silicon re-emerging on physical-AI and real-time on-device demand ref. Compute in memory is usually pitched into that edge and datacentre-accelerator demand, but the source does not name compute-in-memory companies.

Evidence base

Frontier (open questions)

Synthesised 2026-08-31 from 2 KB sources by the resynth pipeline; citations are KB source slugs.

Recent mentions

Frontier questions