100x faster, ultra-efficient photonic processor transforms wireless signal processing and AI at the edge
Engineers at the Massachusetts Institute of Technology (MIT) have unveiled a groundbreaking AI chip that processes wireless signals at the speed of light, promising to revolutionize 6G, edge AI, and real-time signal analysis. Named MAFT-ONN (Multiplicative Analog Frequency Transform Optical Neural Network), the chip performs AI inference directly on raw radio-frequency (RF) signals, without converting them into digital data.
This analog photonic processor achieves up to 100 times the speed of traditional digital AI chips, while drastically reducing power consumption, weight, and size, making it ideal for next-generation cognitive radios, edge devices, and real-time wireless AI.
In lab tests, MAFT-ONN achieved 95% modulation classification accuracy and completed millions of analog multiply-accumulate operations, including image recognition from the MNIST dataset. Thanks to its novel architecture, the chip converts RF data into the frequency domain and processes it entirely optically, eliminating bottlenecks seen in traditional optical neural networks.
“We can fit 10,000 neurons onto a single device and compute the necessary multiplications in a single shot,” said Ronald Davis III, PhD ’24, lead researcher on the project.
Operating near the Shannon limit, the theoretical maximum for information transfer, MAFT-ONN achieves over 85% inference accuracy in just 120 nanoseconds, and can scale up to 99% accuracy with additional measurements, all without sacrificing speed.
Key Advantages:
- 100x faster than conventional digital AI chips
- Ultra-low power and analog-native design
- Compact and cost-effective for widespread deployment
- Real-time AI inference for wireless, medical, and autonomous applications
- Supports emerging 6G and edge AI use cases
“This work is the beginning of something that could be quite impactful,” said Dirk Englund, MIT professor of electrical engineering and senior author of the paper published in Science Advances. “It opens the door to powerful real-time signal processing at the edge, with potential applications from self-driving cars to smart medical implants.”
The team is now working on multiplexing techniques to scale the chip’s capabilities further and adapt it to support transformer-based models and large language models.
