Summary: Researchers have developed the first “microwave brain” chip capable of processing ultrafast data and wireless communication signals simultaneously. Using analog, nonlinear microwave physics instead of traditional digital circuits, the chip can decode radio signals, track radar targets, and classify high-speed data streams in real time.
It offers accuracy comparable to digital neural networks, but with much lower power consumption. This development could transform applications ranging from secure wireless sensing to wearable edge computing.
Important facts:
- Analog Microwave Neural Network: Processes data at tens of gigahertz without a digital clock.
- High performance: The device uses less than 200 milliwatts and has an accuracy of up to 88%.
- Wide applications: From radar registration to advanced AI on wearables.
Source: Cornell University
Researchers at Cornell University have developed an energy-efficient microchip they call a “microwave brain.” It’s the first processor that can calculate both superfast data signals and wireless communication signals using microwave physics.
The processor, published August 11 in the journal Nature Electronics, is the first true microwave neural network and is fully integrated on a silicon microchip.
It performs real-time frequency domain calculations for tasks such as decoding radio signals, tracking radar targets, and processing digital data, all while consuming less than 200 milliwatts of power.
“Because it can generate directly programmable distortion across a wide frequency range, this chip has the potential to handle a wide array of computational tasks,” explained lead researcher Bal Govind, a PhD candidate. He conducted the study alongside Maxwell Anderson, also a PhD candidate. This flexibility sets the technology apart from conventional computing systems, which often rely on rigid hardware configurations and narrowly focused processing abilities.
One of the most significant advantages of this approach is that it bypasses many of the complex signal processing steps typically required in digital computers. By eliminating these extra stages, the system operates with greater efficiency and potentially faster processing speeds, opening doors to real-time applications that previously required substantial computational power and time.
This breakthrough is made possible by designing the chip as a physical neural network—a brain-inspired computing architecture that leverages interconnected nodes embedded in tunable waveguides. These waveguides allow the chip to recognize complex patterns and adapt to new data, effectively enabling it to learn over time. Such a design not only mimics the way the human brain processes information but also represents a step toward more energy-efficient and scalable neuromorphic systems.
But unlike traditional neural networks, which rely on digital operations and step-by-step instructions synchronized with a clock, this network uses analog, nonlinear behavior in a microwave system, allowing it to process data streams at tens of gigahertz. That’s much faster than most digital chips.
“Ball has intentionally moved away from many traditional circuit design approaches to make this breakthrough possible,” explained engineering professor Alyssa Upsell, who co-authored the study with Peter McMahon, associate professor of applied physics and engineering. Rather than sticking to conventional computing architectures, Ball explored new design frontiers to overcome the limitations of existing systems.
Instead of attempting to perfectly replicate the structure of digital neural networks, he developed a system that behaves like a carefully controlled mix of frequency responses. This innovative approach could serve as a foundation for future high-performance computing technologies. By manipulating signal frequencies directly, the chip enables a new kind of analog computation that bridges the gap between digital logic and neural processing.
The chip is capable of performing both basic logic operations and complex computational tasks—such as detecting bit sequences and computing binary values from high-speed data streams. It has demonstrated over 88% accuracy across a range of classification tasks, including identifying different wireless signal types. This performance rivals that of digital neural networks, but with a significantly lower power requirement and a much smaller physical footprint, making it highly promising for compact and energy-efficient AI applications.
“In traditional digital systems, the increasing complexity of tasks requires more circuitry, power, and better error correction to maintain accuracy,” says Govind. “But with our probabilistic approach, we can maintain high accuracy in both simple and complex calculations without this overhead.”
The researchers said the chip’s extreme input sensitivity makes it ideal for hardware security applications, such as detecting anomalies in wireless communications across multiple microwave frequency bands.
“We also believe that by further reducing power consumption, this technology could be applied to edge computing,” said a representative from Apple. “Imagine integrating it directly into a smartwatch or mobile phone, allowing users to build and run local models right on their devices—without needing to constantly rely on cloud-based servers.” This vision aligns with the growing demand for faster, more private, and energy-efficient on-device AI processing.
Although the chip remains in the experimental stage, researchers are optimistic about its long-term scalability and potential for real-world deployment. The architecture’s compact design and low-power operation make it an ideal candidate for devices where space and energy are limited. If fully realized, it could revolutionize how personal electronics process and respond to data in real time.
To move closer to that goal, the team is actively exploring methods to further enhance the chip’s accuracy and ensure compatibility with existing digital and microwave signal processing systems. These ongoing experiments aim to refine the chip’s performance and ease its integration into current technological infrastructures, bridging the gap between emerging analog computing techniques and the digital systems that dominate today’s devices.
Funding: This work is a research project within a larger project, in collaboration with the Defense Advanced Research Projects Agency and the Cornell Center for Nanoscale Science and Technology, and is funded in part by the National Science Foundation.
Abstract
An integrated microwave neural network for broadband computing and communications
The development of bandwidth-intensive applications, such as multi-gigabit communications and radar imaging, requires faster processing.
However, in microwave systems, where the frequency is higher than the clock frequency, sampling and calculation become difficult.
Here we present an integrated microwave neural network for broadband computing and communications.
Our microwave neural network operates at tens of gigahertz, but is reprogrammed with a low bit rate (megabits per second). By exploiting the strong nonlinearity of coupled microwave oscillations, it expresses its calculations in a narrow spectrum, making them easier to read.
The system finds bit sequences in multi-gigabit per second data and emulates digital functions without custom circuitry. It accelerates radio frequency machine learning by classifying coding schemes and detecting frequency shifts to track radar flight paths.
The microwave neural network is fabricated using standard complementary metal-oxide semiconductor technology.
The on-chip wavelength is less than 0.088 mm², and the power consumption is less than 200 mW. This allows integration into a general-purpose analog processor.
