Enlightra Team Demos Photonic AI Engine

A research team led by Enlightra’s founders demonstrated a new method of photonic AI inference that is smaller than earlier photonics implementations. Using photonics also provides opportunities for resource sharing that aren’t available with electronics.

Bryon Moyer

A research team led by Enlightra’s founders demonstrated a new method of photonic AI inference. Borrowing from in-memory compute (IMC) in electronic deep-learning accelerators (DLAs), the approach could reduce the size of a photonic inference engine, but it requires further development and is years from production.

The novel approach combines the startup’s signature laser-generation technology with phase-change-memory (PCM) cells to perform the matrix multiplications that many neural networks require. It configures the memory cells in analog states to act as gates; the amount of transmitted light implements multiplication. The resulting waveforms then mix to perform multiply-accumulate (MAC) functions.

A proof of concept (PoC) achieved two trillion operations per second (TOPS) running at 17mW (or 17fJ/MAC) in a 9x4 matrix, performing 36 MACs in parallel. At more than 100 TOPS per watt, the prototype far exceeds the efficiency of digital DLAs. The company envisions more than tripling performance by reducing various losses and increasing waveguide bandwidth.

Enlightra is a startup developing a comb generator that creates hundreds of tightly spaced laser wavelengths from a single source (see MPR Jul 2022, “Enlightra Prunes Lasers From WDM”); it incorporated after the founders conducted this research. Although the company won’t sell an inference product, the generated carrier signals can implement multiple levels of parallelism and component reuse, making more-efficient use of the engine’s die area. Enlightra’s business interest lies in potential comb-generator sales if this technique achieves uptake.