As more devices connect to the internet and demand faster speeds—especially for things like video calls, cloud gaming, and smart homes—the pressure is on to make wireless networks faster and more efficient.
But with limited bandwidth available, how can we keep up?
Engineers are turning to artificial intelligence (AI) to help manage wireless traffic and reduce delays.
The problem is, most AI systems are too slow and use too much power for real-time wireless communication. Now, researchers at MIT have built a new type of AI processor that solves that problem—by using light.
This new chip, developed by MIT and featured in Science Advances, processes wireless signals at the speed of light.
It can recognize and classify signals in nanoseconds—about 100 times faster than current digital systems. It also uses much less energy, making it ideal for small devices that need to act fast without draining a battery.
The chip is built using photonics, which means it uses light instead of electricity to process data. This allows it to operate much faster than traditional AI chips. It’s also compact, cheap to make, and flexible enough to be used for different tasks, not just wireless signal processing.
This technology could be especially useful for future 6G networks. In 6G, smart devices like “cognitive radios” will need to adapt their signal formats on the fly to keep connections stable and fast. The new chip, called MAFT-ONN (Multiplicative Analog Frequency Transform Optical Neural Network), could help these devices think and react in real-time.
Besides 6G, the chip could also help self-driving cars respond to their environment instantly or allow medical devices like smart pacemakers to monitor a patient’s health without delay.
Most AI chips today process wireless signals by first converting them into images and then running those through deep-learning models. This works, but it’s slow and takes a lot of computing power. Optical systems like MAFT-ONN skip that step. They work entirely in the “frequency domain,” which is where signal data exists before being turned into digital form. This saves time and energy.
Unlike other optical chips, MAFT-ONN only needs one device per layer in a neural network. It also performs both the simple and complex operations required for deep learning using light alone. The researchers used a method called photoelectric multiplication to boost the chip’s efficiency and scale.
In testing, the chip could classify signals with 85% accuracy in just one try, and more than 99% accuracy with a few extra measurements. And it only takes 120 nanoseconds to complete a task—far faster than current digital systems, which work in microseconds.
Lead researcher Ronald Davis III and his team had to build custom machine-learning tools to match the hardware and make use of the chip’s unique physics. Next, they plan to expand the chip’s capabilities, possibly even adapting it for complex AI models like transformers or large language models.
This breakthrough could help reshape not just wireless networks but a wide range of technologies that rely on fast, reliable, and efficient AI.
