The short version, for readers who know decoders but not photonics.
The bottleneck
An error-correcting decoder repeatedly fetches soft information from memory and brings it to logic. At the scales that matter, that shuttling dominates the energy budget — not the comparisons.
The move
Process the log-likelihood ratios in the optical domain. Light does addition and multiplication for free and without the switching energy that gates pay.
The recursive structure matters: the decoder applies the optical computation to soft information, converts back only where a hard decision is genuinely needed, and feeds that decision forward into the next iteration. Analog circuitry is used for decision-making alone.
LLR ──▶ [optical domain] ──▶ decision ──┐
▲ │
└───── feedback ────────────┘
The result
At least five orders of magnitude better energy efficiency than the electronic counterpart. More interesting than the number is the scaling: power grows linearly with codeword length, not with computational complexity. That is a different kind of hardware, not a faster version of the same one.
What it does not solve
The electro-optical conversion count is still the thing to watch. Minimising conversions was where most of the design effort went, and it remains the obvious target for anyone picking this up next.
- optoelectronics
- compute-in-memory
- ecc