Analog Computing

Explainer

Analog Computing explained

Analog computing represents numbers as continuous physical quantities, such as voltage or current, rather than as strings of 0s and 1s. The physics of the circuit does the calculation. For example, Ohm's law and Kirchhoff's current law let a grid of resistors multiply and add numbers at once.

This suits the multiply-accumulate operations at the heart of neural networks. In an analog AI chip, the weights of a network are stored as conductances in a memory array, an input voltage goes in, and the summed current out is the result. Analog circuits were common before digital computers, and the idea is being revived for AI.

How it differs from the usual approach

A digital chip gives repeatable results by representing every value with a fixed number of bits. An analog chip is approximate: noise, temperature and manufacturing variation shift the values, so precision is typically limited to roughly 4 to 8 bits. In return it can use far less energy for the matrix operations in neural networks, because it avoids moving data and converting every step. Converters between analog and digital signals can take a large share of the power.

Where it is used

• Low-power AI inference in always-on devices and sensors.
• Keyword spotting and simple vision tasks at the edge.
• Analog front ends in radios, audio and sensing circuits.
• Research accelerators for neural networks and optimisation.
• Specialised solvers for differential equations in research.

What to watch

• Analog AI chips have mostly been demonstrated in research or small-volume products, and larger production use is still limited.
• Accuracy under noise and drift is the main engineering problem, handled by training networks to tolerate error.
• Memory technologies such as flash, ReRAM and phase-change memory are candidates for storing analog weights.
• Whether the energy saving survives once converters and control circuits are counted decides if the approach scales.

Frequently asked questions