Direct tensor processing with coherent light

1. https://www.nature.com/articles/s41566-025-01799-7

These require a special chip: Thin-Film Lithium Niobate (TFLN) chips are the optical equivalent of the silicon chips in your computer, but instead of moving electrons (electricity), they move photons (light). Lithium Niobate is a crystal that changes its optical properties when exposed to electricity (the Pockels effect). "Thin-Film" means engineers have figured out how to slice this crystal into extremely thin layers (less than 1 micrometer) and paste it onto a silicon wafer.

For the POMMM system to work, the chips must be 2D, and this is the biggest engineering hurdle right now. 1D vs. 2D: 1D (Current Tech): Most TFLN chips sold today are "linear modulators." They have one input and one output, like a single fiber optic cable. They process a single stream of data (Scalar). 2D (AI Requirement): Neural networks process Matrices (grids of numbers). The POMMM prototype works by flashing a whole "image" of data (e.g., a 100×100 grid) at once. To process this instantly, the chip needs a 2D grid of pixels (100×100 modulators) working simultaneously. The Engineering Problem: It is easy to put wires on a 1D line. It is incredibly hard to wire a 2D grid of optical pixels because the wires for the center pixels get in the way of the outer pixels. This is why 2D TFLN arrays are currently stuck in research labs.

Speed Factors

1. Grid Size N×N
2. Clock Frequency (Cycle Speed)
3. I/O Device Speed (Modulation/Detection)
4. Multiplexing (Wavelength/Polarization)