Contact

← Projects · Oct 2026

Etalon

An experiment in photonic simulation on your own GPU: FDTD for waveguides, bends and resonators, with every CUDA kernel checked against a NumPy reference and no design data leaving your machine.

etalon-studio.netlify.app ↗

Etalon simulates passive photonic devices, such as waveguides, bends, couplers and ring resonators, with the finite-difference time-domain (FDTD) method, on one NVIDIA GPU on your own desk. You describe the device in Python, and Etalon Studio, a local web app in English and Simplified Chinese, shows the setup, the run and the fields. There is no cloud and no telemetry, so design data never leaves the machine.

Correctness is the point. Every CUDA kernel has a NumPy reference it must match, and on the first milestone, a 2D solver, the two agree bit for bit in single and double precision. The measured phase velocity stays within 1.3 × 10⁻⁴ of the Yee dispersion relation, the absorbing boundary reflects at −91 dB, and a 2000 × 2000 grid runs at 2.9 × 10⁹ cell updates a second on an RTX 4060 Ti. A benchmark harness compares it with Meep on five reference devices.

I built it alone in October 2026, as an experiment in how far one consumer GPU and strict numerical checks can go. It is 2D only for now: 3D, a mode solver and S-parameters are not there yet.

Etalon’s demo page: a Bragg mirror simulation, a point source in front of eight silicon layers in oxide, with the reflected light interfering with the incoming wave; a list of recorded runs on the left, and the run parameters and a field profile on the right.
The demo page replays fields the solver computed: here, light reflecting off eight silicon layers.
Etalon Studio’s Results tab for a strip waveguide: the Hz field guided along the waveguide at 1.55 µm, the run’s GPU, grid and timing above it, and the monitor, component and wavelength controls on the right.
Etalon Studio: a strip waveguide’s field at 1.55 µm, from a three-second run on an RTX 4060 Ti.