For decades, the most elusive goal in artificial intelligence has been to reproduce one of the brain’s defining tricks: computing where information is stored instead of constantly shuttling data between separate memory and processing units. A new photonic computing architecture called FLARE brings that idea into a large-scale system that combines light-based computation with electronic memory. The researchers behind the system report a monolithic chip containing 7,378 artificial neurons, capable of retaining information for seconds while also responding with the rapid dynamics required for high-speed processing. In a demonstration involving autonomous racing-drone navigation, a ten-core version of the system performed sensing, exploration and adaptation with a reported system-level energy cost of just 61.87 attojoules per operation. The work points toward machines that could sense and interpret their surroundings with far less dependence on conventional digital hardware.
To read more, click here.