Post-Digital FuturesSignals · systems · strategyWork with Dean

Post-Digital Primer 05

In-memory and analog computing

The opportunity is to reduce data movement and exploit device physics for selected operations. The penalty is precision, variability and programmability.

Why it matters

Moving data between memory and processors is a major energy and latency cost in AI and signal processing. Computational memory could reduce that burden in edge equipment and datacentres if converters, variability and software do not erase the gain.

What it is

In-memory computing performs operations inside or next to storage arrays instead of repeatedly moving data to a processor. Analog computing represents quantities continuously and lets voltages, currents, charge or waves perform mathematical operations directly.

What it is not

The categories overlap but are not identical. A digital near-memory accelerator may avoid analog uncertainty; an analog RF computer may have no memory array. Device-level TOPS/W figures often omit converters, control and error correction.

Telecoms read-acrossData movement is a genuine energy bottleneck. Credible systems use analog or in-memory blocks for stable, repeated calculations while conventional processors retain control, precision and software compatibility.

How it works

01Store weights

Store model settings as electrical conductance, charge or magnetic states inside an array that can also perform calculations.

02Apply inputs

Encode data as voltages, currents or RF signals across rows or channels.

03Accumulate

The array performs many multiplications and additions in parallel, which are common operations in AI models.

04Correct

Converters and digital logic calibrate and handle operations requiring greater precision.

Use cases

Potential uses.

Status labels show use-case maturity.

Edge neural inference

Run convolution and matrix operations in gateways, cameras and industrial devices.

Advanced prototype

Sensor preprocessing

Extract compact features before data reaches a conventional processor or network.

Plausible near term

RF-domain compute

Use mixers, propagation or antenna arrays to perform portions of inference.

Research

Datacentre acceleration

Use specialist arrays for repeated workloads where the setup cost can be spread across many calculations.

Early product path

Telecom relevance

Possible telecom roles.

Edge gateways

Accelerate stable models while retaining a general processor for updates and unsupported operations.

RAN and RF vendors

Investigate estimation and classification but isolate experiments from deterministic loops.

Cloud operators

Evaluate memory bandwidth, update frequency and utilisation before assuming lower cost.

Chip suppliers

Design around converter energy, endurance, calibration and software.

Reality check

Pre-requisites and proof-points needed

These questions test whether the technology can become a practical, supportable system.

  1. 01What accuracy and precision are sustained?
  2. 02Are converters and control energy included?
  3. 03How often must weights be rewritten?
  4. 04Does device variation reduce yield or portability?
Related signal coverageSignals 016–018 and 034
Browse the signal archive →Back to all explainers ↑