[ALT] Analog Computing

Analog computers solve differential equations and perform continuous mathematical operations by direct physical analogy — voltages representing variables, capacitors representing integrals — rather than by discrete numerical approximation. They were real competitors to digital computers through the 1960s and remain unsurpassed for specific ultra-low-latency or energy-constrained applications.

The Fork

What won: Digital computing — discrete, programmable, general-purpose, and exponentially improving via Moore's Law. By the mid-1970s, digital computers had surpassed analog for all routine scientific computing.

What was abandoned: Analog computers — differential analysers (Bush 1931), electronic analog computers (1940s–60s), and analogue simulation systems. Used for flight simulation, missile trajectory calculation, power grid control, and nuclear reactor modelling. They could solve differential equations faster than real time by simply running the physical system, without the discretisation errors of numerical methods.

Why Analog Computers Were Superior for Certain Tasks

• Speed: An analog circuit's output changes at the speed of electrons through a resistor-capacitor network — effectively instantaneous for signal bandwidths up to MHz or GHz, far faster than any numerical solver.
• Energy efficiency: A continuous-time signal processor can perform multiply-accumulate operations using only thermal noise-floor power — orders of magnitude more efficient than digital switching.
• No discretisation error: Exact solution of differential equations (within component tolerances) rather than numerical approximation.

The key weaknesses: poor programmability (hardware must be rewired for new problems), limited precision (component tolerances dominate above ~4 decimal digits), and no memory for general-purpose applications.

Current Status

Niche revival — analog computing is experiencing a renaissance driven by AI inference energy costs. IBM, Intel, and MIT spinoffs are building mixed-signal neural network accelerators that perform matrix multiplication in the analog domain (using memristors, phase-change materials, or current-mode circuits) then digitise only the output — achieving 100–1000× better energy efficiency for inference than digital multipliers. Mythic AI (Texas), Aspinity, and Analog Inference are commercialising this approach. The neuromorphic computing movement is its intellectual heir.

Discovery Character

Surprise level: Low for the original abandonment (digital's programmability advantage was clear) / High for the revival (the AI energy crisis has made analog competitive again in ways 1970s engineers did not predict).

Mode: Systematic engineering of both paradigms; the fork was determined by Moore's Law and programmability requirements, not by analog's intrinsic limits.