MIT researchers develop photonic processor for 6G wireless signal processing

MIT researchers announced a new photonic AI hardware accelerator, the Multiplicative Analog Frequency Transform Optical Neural Network (MAFT-ONN), designed for wireless signal processing. The chip performs machine-learning computations using light, processing wireless signals in approximately 120 nanoseconds, about 100 times faster than leading digital alternatives. It achieves 85 percent accuracy in signal classification in a single shot, converging to over 99 percent with multiple measurements.
The MAFT-ONN operates in the frequency domain before signal digitization, performing all linear and nonlinear operations in-line, requiring only one device per neural network layer. This design, utilizing photoelectric multiplication, enables scalability by fitting up to 10,000 neurons on a single device for computations in a single shot. The chip is smaller, lighter, cheaper, and more energy-efficient than digital AI accelerators, making it suitable for edge devices in applications like cognitive radios for 6G networks, autonomous vehicles, and smart pacemakers.
The researchers faced challenges in mapping machine-learning computations to optical hardware, requiring a customized framework to leverage the physics of light for efficient processing. The MAFT-ONN processes wireless signals to classify modulation types, enabling devices to infer and extract data. Future work includ...