Modern AI inference is bottlenecked by the energy cost of multiply-accumulate (MAC) operations performed on digital, transistor-based hardware. As AI models grow, so does the power draw of the hardware needed to run them — a barrier that keeps advanced inference out of reach for cost- and power-constrained users, from students to developing-world engineers. This project asks a different question: what if the core arithmetic of a neural network could be done by light itself, using parts already sitting in millions of discarded calculators?

The Photonic TPU is a proof-of-concept optical MAC cell. Each input value modulates an LED's brightness via pulse-width modulation. The resulting beam passes through a liquid-crystal (LC) cell whose transmittance — governed by Malus's Law (I=Iₒcosϴ) — encodes a trained weight. Two independently modulated beams converge on a single photodiode, which sums their intensities by physics: because the beams are mutually incoherent, no interference or path-length matching is required, and the addition step consumes no clock cycles and no adder circuit. This 2-input, 1-output cell is the atomic unit of the architecture; a full neural network layer is simply M copies of this cell operating in parallel, one per output neuron, each receiving all N input beams simultaneously (see accompanying MAC-cell diagram and neural-network layer-tiling diagram).

Prior optical MAC research — including free-space matrix-vector multiplication work using offset and scaling to convert signed-weight computation into non-negative optical intensities, avoiding balanced photodiode pairs (Wang et al., 2021, 2022) — has demonstrated this physics works, but almost exclusively on cleanroom-fabricated silicon photonic chips costing thousands of dollars and requiring months of fabrication lead time. This project's novelty is not the physics, which is established, but the substrate: every optical and electronic component — LC weight cell, photodiode, and driving electronics — is salvaged from obsolete consumer calculators rather than purpose-built or custom-fabricated. This makes the core contribution one of accessibility: proving that the fundamental operation of optical AI acceleration can be replicated for the cost of a used calculator, rather than the cost of a semiconductor fab run.

The benefits follow directly from this: near-zero marginal power cost for the addition step, since it is physical rather than computed, and a bill of materials cheap and simple enough to be replicated in classrooms, hobbyist labs, and resource-constrained research settings worldwide — not just institutions with cleanroom access. The natural application space is low-power edge inference and, more broadly, AI-hardware education: giving students and independent researchers a physical, buildable entry point into a field currently gated by expensive infrastructure. As a manufactured product, the underlying components — LC panels and photodiodes — are already produced at massive scale and low cost for the display and calculator industries, suggesting a purpose-built version of this design could undercut both digital accelerators and research-grade photonic chips on cost, while opening an entirely new, low-power computing category to markets currently priced out of AI hardware altogether.

Feasibility rests on the fact that every element of this design is grounded in established, individually-verified physics rather than speculative engineering.