Cognifiber Targets Purely Photonic AI

The Israeli startup has built a proof-of-concept system for a simple optical neural network, but it will require additional effort to handle commercial AI models.

Linley Gwennap

Let there be light! Startup Cognifiber has an ambitious goal to create a new, more efficient computing method using light instead of electricity. Several other companies are also developing photonic computing, with a particular focus on deep neural networks, but their efforts have been hampered by the need to pair optical accelerators with digital processors. Cognifiber hopes to perform AI entirely using photonics, eliminating the overhead of converting between the optical and digital domains.

CEO Eyal Cohen and CTO Zeev Zalevsky founded Cognifiber in 2016, after Cohen had developed and published the basic concepts of his photonic computing while doing postdoctoral research under Professor Zalevsky at Israel’s Bar-Ilan University. The company remained a side hustle until it received $2.5 million in 2019 from Chartered Group, a Singaporean financial firm that invested another $6 million last year.

With this funding, Cognifiber has grown to 10 full-time employees and another dozen or so part-timers. The Israel-based team has built two proof-of-concept designs to demonstrate its technology. The startup plans to debut its first product later this year.

Cognifiber has developed a proof-of-concept system it calls the MVP. It features 28 neurons, each with its own photodetector and laser. The system can take 16 analog input signals and implement a four-layer network having up 96 weights. At 100MHz, it produces 100 million inferences per second. Each inference, however, is limited to networks that fit into 28 nodes. The MVP consumes 300W and performs 19 billion operations per second (GOPS).