The oscillator AI chip

Most sensor data is thrown away unread. A microphone, a strain gauge or a vibration pickup produces a continuous signal around the clock, and the machinery that decides whether anything in it matters — an analog-to-digital converter, a processor, a Fourier transform — burns most of a device's power budget doing arithmetic on silence. The oscillator chip attacks that from the physics side: let the signal run through a small bank of damped harmonic oscillators, each one two analog integrators in a loop, and read the bank out through integrators that average on exactly the time scale a decision needs. Filtering, temporal accumulation and a bounded decision then happen by virtue of the circuit obeying its own differential equation, with no clock and no conversion until something is worth waking up for.

The hard questions are not whether such a network can classify — it can, and the published work on harmonic oscillator recurrent networks shows it — but what survives contact with the machine. An analog coefficient lives on a twelve-bit converter inside a fixed range. An integrator output is clamped by a diode limiter that stops short of the nominal rail, at a level that was measured, not assumed. A node's resonance frequency is set jointly by two programmable coefficients and two discrete time-scale factors. Whether a network trained in floating point still works once every number is forced onto that grid is a question that deserves measurement, not argument, and it is the question the digital twin behind this page was built to answer.

Behind the glass: a live, element-faithful model of the architecture you can play with in the browser, and the twin's measured results on what training inside the machine's range costs.