With the rapid increase in network devices, applications such as video conferences, cloud games and smart homes are increasing in demand for bandwidth, and limited management of wireless spectrum has become a major challenge at the moment. The key to solving this problem is high-performance real-time processing technology, and the latest breakthrough of the MIT research team is expected to change the situation in a complete way.
The optical AI processor developed by MIT has been published in the journal Science Advanceds, which can process wireless signals at the speed of light, achieving unprecedented computing efficiency. This chip, called MAFT-ONN (multiplication frequency conversion optical neural network), can complete signal classification in nanoseconds, with a speed of nearly 100 times faster than existing digital systems, and its energy consumption is much lower than that of traditional AI hardware.
Unlike traditional digital processors that need to convert the signal to an image first and then process it, MAFT-ONN operates directly in the domain, skipping the digital conversion step, achieving significant energy and acceleration. The measurement of this chip can reach 85% classification accuracy, and the measurement of multiple times can exceed 99%. The entire process only takes 120 nanoseconds, which is far away from existing technology.
MAFT-ONN is uniquely designed. Each neural network layer requires only one device, which can integrate 10,000 neural cells into a single chip, and adopt "optical multiplication" technology to achieve high-efficiency computing. The research team must develop the machine learning architecture on its own to match the hardware characteristics to ensure that the physical characteristics of the chip are fully utilized.
This technology is particularly important in the future 6G network, and can make smart devices such as "cognitive wireless" adapt to network conditions in real time and maintain stable connections. Research team member Ronald Davis III said that as the measurement time increases, the accuracy will increase, and MAFT-ONN's ultra-fast processing time makes this process not affecting performance at all.
Dirk Englund, a professor in the Department of MIT Motors and Computer Science, pointed out: "This technology is a real-time and reliable AI recommendation to unlock many possibilities and is an open end with deep impact." This chip can not only change wireless communication, but can also be used in real-time environmental reactions in self-driving cars, smart medical devices and other fields.
The research team plans to use "copy technology" to improve chip efficiency and expand to a more complex deep learning architecture, including Transformer models and large language models. This breakthrough will deeply affect various technical fields that require real-time and efficient AI, laying the foundation for the true era of smart connections.
MIT’s Optical AI Chip That Could Revolutionize 6G at the Speed of Light Light-speed AI chip could power the future of 6G wireless