Analog Neural Network

last updated 2026-08-31

Physics / mechanism

An analog neural network implements the arithmetic of a neural network using continuously varying physical quantities rather than binary digits. The framing offered in the available source is that digitisation is a modelling choice rather than a physical necessity: the world is continuous, and binary representation is convenient but not intrinsically matched to the signals being processed ^ref.

The supplied material does not specify device physics, precision, energy figures, or circuit topologies for analog neural network accelerators. What it does identify is the commercial form factor expected to carry the approach: mixed-signal integrated circuits that combine analog and digital blocks on the same die, targeted at edge AI inference ^ref.

Key parameters that would determine viability (bit-equivalent precision, drift and temperature stability, analog-to-digital conversion overhead, area per multiply-accumulate) are not covered by the source and should be treated as open.

Competitive landscape

The source places analog computing alongside several other non-conventional compute approaches covered in the same series: neuromorphic computing, optical computing, and, prospectively, quantum, mechanical, molecular, magnetic and acoustic computing ^ref. It treats these as distinct entries rather than a single category, but does not give a technical comparison between them. Analog neural networks therefore sit in this reference base as one member of a cluster of alternatives to digital CMOS von Neumann inference, with the mixed-signal variant presented as the nearest-term commercial expression.

Evidence base

The stated thesis is that mixed-signal analog-digital ICs capture 50% of a $60bn edge AI hardware market, framed as an underrated outcome ^ref.
Analog computing is presented as one of a sequence of exotic computing paradigms alongside neuromorphic and optical computing, published 23 July 2023 ^ref.
The motivating argument is representational: continuous physical signals are reduced to binary for convenience rather than fidelity ^ref.
The Antikythera Mechanism is cited as the oldest analog computer, establishing analog computation as a pre-digital lineage rather than a new invention ^ref.

Frontier (open questions)

What effective bit precision do current analog neural network accelerators achieve on standard inference benchmarks, and does it hold across process, voltage and temperature variation?
What fraction of total energy in a mixed-signal inference chip is consumed by data conversion (ADC/DAC) rather than by the analog compute array itself?
Is the $60bn edge AI hardware market figure sourced from a specific forecast, and what year does the 50% mixed-signal share refer to?
Which workloads (keyword spotting, vision, transformer inference) have been demonstrated end-to-end in analog silicon, with measured accuracy against a digital baseline?

Synthesised 2026-08-31 from 1 KB sources by the resynth pipeline; citations are KB source slugs.

Recent mentions

2023-07-23 🔮 E09: Analog Computing: The Once and Future King substack

Frontier questions

What effective bit precision do current analog neural network accelerators achieve on standard inference benchmarks, and does it hold across process, voltage and temperature variation?
What fraction of total energy in a mixed-signal inference chip is consumed by data conversion (ADC/DAC) rather than by the analog compute array itself?
Is the $60bn edge AI hardware market figure sourced from a specific forecast, and what year does the 50% mixed-signal share refer to?
Which workloads (keyword spotting, vision, transformer inference) have been demonstrated end-to-end in analog silicon, with measured accuracy against a digital baseline?