Photonic memory is the attempt to hold information in, or addressably alongside, light rather than in charge on a transistor, and it matters because the absence of a good one is the standing reason photonic computers still have to convert back to electronics.
Summary
Light is an excellent carrier of information and a terrible container for it. A photon in a waveguide does not sit still, so any “photonic memory” has to do one of four quite different things: recirculate the light (delay lines, mirror cavities), park it in an atomic ensemble and retrieve it later, freeze it into a material state that light can subsequently read (phase-change alloys, implanted colour centres, magnetic domains), or exploit nonlinear feedback so an optical cavity settles into one of several stable states. The label covers all four, and they share almost nothing but the word.
The benchmark to beat is unforgiving. HBM3E already delivers around 1.2 TB/s per stack at roughly 3-4 pJ/bit, HBM4 doubles the bus to 2,048 bits for 2.0-3.3 TB/s, and the market was around $35B in 2025 heading for a claimed $100B TAM by 2028. Nothing in the photonic memory literature is currently quoted against those numbers, which is itself informative.
Evidence base
Open questions
- Does a fabricated Sb2Se3 programmable mode converter array reproduce the simulated 5-bit, 32-level programming, and what is the measured insertion loss per cell and the resulting maximum array depth?
- What are the write energy, write throughput and endurance of any photonic memory cell, expressed per bit, so it can be set against the 3-4 pJ/bit of HBM?
- Can the EBL-plus-ion-implantation storage medium be rewritten, or is its addressable market permanently cold archival?
- Does any foundry offer a qualified PCM photonic memory cell in a PIC process design kit, and on what schedule?
Assessment drafted 2026-08-31 from up to 16 KB sources using the technology-scorecard framework; scores are a draft read pending review.