Constructing the Computer Revolution

Computer Engineering is designing the hardware and software behind next generation computing

While tools such as the large language model-based chatbot ChatGPT have laser focused our collective attention on artificial intelligence, most of us have been using AI every day, for longer than we realize: for unlocking a phone with Face ID, conducting a search online, or relying on Google maps to get us to our destination. The use of smart technology has continued to explode ­– we can tell Google Home to order us a pizza or Alexa to turn on the lights.

This increased demand for smaller, faster, and more secure hardware needed for modern computing has transformed the field of computer engineering. Chips have become incredibly powerful, but this alone simply isn’t enough to meet modern needs. Designing hardware with the software it will run, and utilizing software to improve hardware design enables optimized performance, increased efficiency, and reduced design time and cost.

Today’s supercomputers and high-performance computing applications can have billions or even trillions of transistors, so it’s no longer possible to connect and test them manually. Now, much of this work is done automatically by software developed by computer engineers, through electronic design automation processes.

“From the moment that we get specifications of what we want our system to be doing, how it should be performing, and what it should look like, all the way to when we get a piece of hardware, 80% of this work is done automatically,” said Assistant Professor Inna Partin-Vaisband. “There is a lot of computer science labor for this kind of research, but that is applied to creating new hardware, optimizing it, and making it safe and secure.”

The ability to cull unnecessary information can speed processing, which is essential with the onslaught of data generated by AI applications.

Navigating a data deluge

Partin-Vaisband and Associate Professor Amit Ranjan Trivedi were recently awarded a grant as part of a project to develop sensors that mimic the brain’s ability to focus on what’s important. Electronic sensors can perceive or “see” everything around them, generating too much information to be stored or processed. The duo will address challenges related to real-time trust and sensor reputation tracking.

Partin-Vaisband is also developing an unsupervised machine learning framework to detect hardware Trojans, malicious tiny circuits that can be embedded in chips, and can interfere with the chip’s operation, steal information, and modify data that’s being sent. These infected circuits are rarely triggered and are difficult to find.

“Making our chips more secure is a high priority in the CHIPS and Science Act and will become more and more important as chips continue expanding into security-sensitive domains such as personalized medicine, artificial intelligence, autonomous vehicles, and others,” Partin-Vaisband said.

Fortifying, and improving hardware designs

A team of engineers led by Associate Professor Pai-Yen Chen and Professor Ahmet Enis Cetin is securing the communication between wireless devices using quantum physics-inspired remote monitoring. By incorporating a mathematical model that pinpoints what is known as exceptional points – or the least likely result from within a range of numbers – random, unique digital fingerprints can be generated. These unique fingerprints, generated from radiofrequency and analog integrated circuits, are known as Physical Unclonable Functions (PUFs) and can be used with both radio wireless devices such as RFIDs, near-field communication tags, and IoT devices, making them far safer.

Chen said the concept for using these exceptional points was inspired by the variations in the human vocal tract that led to billions of different voices; yet humans and machine learning algorithms can recognize each person by the uniqueness of their voice. Like with human voices, the sample-specific uniqueness of particular wireless devices can be derived from the mixed amplitude, frequency, phase, and polarization features of radio waves sourced or scattered from an antenna, circuitry, or surface.

Trivedi and Cetin are part of a team designing computer chips that will allow for faster, more precise processing of machine learning and artificial intelligence applications, by using a compute-in-memory approach, where processor and storage elements are intertwined, eliminating the need for an external processor. This will save costs and more importantly, cut down on processing time by eliminating excessive data exchanges. This is key for applications that require calculations to be made in real time, such as robotic surgery or piloting a drone.

Semiconductors can also aid with the new constraints to size, power, and reliability posed by AI. Trivedi is studying new semiconductor materials with a team of researchers from multiple disciplines that can work with very complex, random numbers, which are in line with the fundamental nature of reasoning and AI software, rather than traditional binary on-off switches.

Enabling new scientific tools

Trivedi hopes improved semiconductors can help another area of study take flight. He works with tiny drones—think the size of insects—that will use machine learning to self-navigate. Using electronic sensors or lightweight cameras, for example, a network of these little drones could perform tasks such as surveying confined industrial spaces to quickly identify sources of poisonous gas leaks, passing underneath a door or through debris to find survivors during a search and rescue operation, or locating infected plants in an agricultural field to stem the spread of disease. Due to the diminutive size of these drones, limiting the number of switches is essential.

“If we can implement these computations using new kinds of semiconductors, we might have a much better shot of achieving this vision of insect-scale autonomous robots,” Trivedi said.

Cetin has developed AI-enabled sensors that can detect wildfires, a task that previously required hands-on human intervention. Using visual odometry and drones, the position and orientation of a wildfire can be determined by analyzing ordinary camera images via a user-friendly application.

“Deep neural network-based AI methods enable a single live camera to monitor a vast area from a distance, and detect wildfire smoke in real time,” Cetin said. “We are developing computationally efficient algorithms that can be implemented inside the camera and sensors, replacing energy-consuming operations.”

The future of hardware design

As computing broadens and specialized applications continue to grow, creating new hardware, optimizing it, and improving security will increasingly rely on hardware-software co-design.

“This is a new flavor and new energy in the computer engineering area that we are witnessing,” Trivedi said, “It will play a very important role in defining how next-generation hardware will be built.”