MIT sees astonishing light-speed 6G processing with its new "100 times faster" chip

Data traffic is growing at an almost dizzying pace. Edholm’s law shows that both data rates and spectrum needs keep rising exponentially. At the same time, deep neural networks are gobbling up more computing power just as Moore’s law is slowing. This mismatch has pushed engineers to hunt for new ways to handle future networks like 6G.

A team at MIT (Massachusetts Institute of Technology) built a new AI chip designed just for wireless signals. Their device is called a multiplicative analog frequency transform optical neural network (MAFT-ONN). It works entirely in analog form on raw radio-frequency (RF) signals. In lab tests it handled modulation classification that quickly reached 95 percent accuracy. It also ran nearly four million fully analog multiply-accumulate operations to recognize handwritten digits from the MNIST (Modified National Institute of Standards and Technology) dataset.

Classical optical neural networks often hit roadblocks on scaling up and end up with a lot of extra hardware. MAFT-ONN solves this by converting signals into the frequency domain before any digitizing happens. Every layer uses a single optical processor to do both the straight-line (linear) and the more complex (nonlinear) math right on the spot. “We can fit 10,000 neurons onto a single device and compute the necessary multiplications in a single shot,” says Ronald Davis III PhD ’24.

Thanks to moving data in analog form near the Shannon capacity limit (which defines the maximum amount of information that can be transmitted over a communication channel), MAFT-ONN runs hundreds of times faster than regular RF receivers. In a single 120-nanosecond shot it hit 85 percent accuracy. By taking a few more measurements it can climb above 99 percent. “The longer you measure, the higher accuracy you will get. Because MAFT-ONN computes inferences in nanoseconds, you don’t lose much speed to gain more accuracy,” Davis adds.

Compared to digital AI chips, this photonic (since it's based on light) processor is about 100 times faster while sipping far less power. It’s also smaller, lighter, and cheaper. That makes it a natural fit for edge gadgets like cognitive radios that tweak their modulation formats in real time to push data rates higher and cut interference.

“There are many applications that would be enabled by edge devices that are capable of analyzing wireless signals. What we’ve presented in our paper could open up many possibilities for real-time and reliable AI inference. This work is the beginning of something that could be quite impactful,” says Dirk Englund, professor of electrical engineering and computer science at MIT and senior author of the paper in Science Advances.

Pushing deep-learning at light speed could help beyond wireless. It might let self-driving cars react in the blink of an eye or allow smart pacemakers to keep a constant watch on heart health. Next, the team plans to add multiplexing schemes to boost computation even more and to adapt the design for larger AI models like transformers and large language models.

Source: MIT News, Science Advances

This article was generated with some help from AI and reviewed by an editor. Under Section 107 of the Copyright Act 1976, this material is used for the purpose of news reporting. Fair use is a use permitted by copyright statute that might otherwise be infringing.

Source

Hope you enjoyed this news post.

Posted Tuesday 5 August 2025 at 3:47 am AEST (my time).

News posts... 2023: 5,800+ | 2024: 5,700+ | 2025 (till end of July): 3,458

RIP Matrix | Farewell my friend