The optical AI chip reads wireless signals fast, using light and simple steps. It works fast, saves energy, and helps in real-time processing.

Credits:Credit: Sampson Wilcox, Research Laboratory of Electronics
MIT researchers have developed an AI hardware accelerator for wireless signal processing. It uses an optical processor that performs machine-learning tasks at the speed of light. The chip uses a multiplicative analog frequency transform optical neural network (MAFT-ONN), which processes signals directly in the frequency domain before digital conversion. It performs both linear and nonlinear operations in-line for deep learning.
Unlike systems that need one device per neuron, MAFT-ONN uses one device per network layer, fitting 10,000 neurons on one device. Photoelectric multiplication completes multiplications in one step, boosting efficiency and simplifying scaling. The system is smaller, cheaper, uses less energy, and classifies wireless signals in nanoseconds with about 95% accuracy.
This optical system processes signals about 100 times faster than digital alternatives. Its accuracy allows quick and reliable classification. Keeping all operations in the frequency domain reduces complexity and hardware needs.
With one device per network layer, MAFT-ONN lowers hardware demands and makes scaling easier. Photoelectric multiplication increases efficiency by completing computations in one step. The machine-learning framework matches the behavior of the optical hardware to fully use its capabilities.
The system recognizes modulation types in wireless signals, helping edge devices decode data automatically. Tests showed MAFT-ONN reached 85% accuracy in one measurement and over 99% with multiple short measurements, all within 120 nanoseconds. This is much faster than digital radio frequency devices working in microseconds.
This technology enables real-time processing for applications. Autonomous vehicles can make decisions as conditions change. Healthcare devices like pacemakers can monitor and respond to heart conditions. 6G wireless networks can benefit as cognitive radios adapt data rates to changing environments.
Because of its design, MAFT-ONN can be adopted without major infrastructure changes, lowering costs and reducing energy use.
Researchers plan to expand MAFT-ONN with multiplexing schemes for more computations and apply it to deep learning models like transformers and large language models, extending its use beyond wireless signal processing.