Research Bits: Oct. 18

Modular AI chip

Engineers at the Massachusetts Institute of Technology (MIT), Harvard University, Stanford University, Lawrence Berkeley National Laboratory, Korea Institute of Science and Technology, and Tsinghua University created a modular approach to building stackable, reconfigurable AI chips.

The design comprises alternating layers of sensing and processing elements, along with LEDs that allow the chip’s layers to communicate optically.

“You can add as many computing layers and sensors as you want, such as for light, pressure, and even smell,” said Jihoon Kang, a postdoc at MIT. “We call this a LEGO-like reconfigurable AI chip because it has unlimited expandability depending on the combination of layers.”

“As we enter the era of the internet of things based on sensor networks, demand for multifunctioning edge-computing devices will expand dramatically,” added Jeehwan Kim, associate professor of mechanical engineering at MIT. “Our proposed hardware architecture will provide high versatility of edge computing in the future.”

The team’s first design was configured for basic image recognition tasks, layering image sensors, LEDs, and arrays of memristors that can be trained to process and classify signals directly on the chip.

“The team’s optical communication system consists of paired photodetectors and LEDs, each patterned with tiny pixels. Photodetectors constitute an image sensor for receiving data, and LEDs to transmit data to the next layer. As a signal (for instance an image of a letter) reaches the image sensor, the image’s light pattern encodes a certain configuration of LED pixels, which in turn stimulates another layer of photodetectors, along with an artificial synapse array, which classifies the signal based on the pattern and strength of the incoming LED light,” said Jennifer Chu of the MIT news office.

“Other chips are physically wired through metal, which makes them hard to rewire and redesign, so you’d need to make a new chip if you wanted to add any new function,” said Hyunseok Kim, a postdoc at MIT. “We replaced that physical wire connection with an optical communication system, which gives us the freedom to stack and add chips the way we want.”

The fabricated chip measured about 4mm² and was stacked with three image recognition blocks tuned to classify one of three letters, M, I, or T. The researchers then shone a pixellated image of random letters onto the chip and measured the electrical current that each neural network array produced in response. The chip was able to correctly classify clear images, but had trouble with blurry ones. By swapping the processing layer for one with better denoising, accuracy improved. “We showed stackability, replaceability, and the ability to insert a new function into the chip,” said Min-Kyu Song, a postdoc at MIT.

Next, the researchers plan to develop additional sensing and processing capabilities for the chip.

Vertical ferroelectric and antiferroelectric FETs for data storage

Researchers at the University of Tokyo developed a proof-of-concept 3D stacked memory cell for data storage based on ferroelectric and antiferroelectric field-effect transistors (FETs) with atomic-layer-deposited oxide semiconductor channel.

The memory is made of hafnium oxide and indium oxide layers that are deposited in a vertical trench structure. The FETs store data in a non-volatile manner, while the vertical device structure increases information density and reduces operation energy needs.

“Ferroelectric materials have electric dipoles that are most stable when aligned in the same direction. Ferroelectric hafnium oxide spontaneously enables the vertical alignment of the dipoles. Information is stored by the degree of polarization in the ferroelectric layer, which can be read by the system owing to changes in electrical resistance. On the other hand, antiferroelectrics like to alternate the dipoles up and down in the erased state, which enables efficient erasure operations within the oxide semiconductor channel,” the researchers explained.

“We showed that our device was stable for at least 1,000 cycles,” said Zhuo Li of the University of Tokyo.

The team experimented with various thicknesses for the indium oxide layer. Optimizing this parameter can lead to significant increases in performance, said Masaharu Kobayashi, an associate professor in the Institute of Industrial Science at the University of Tokyo. “Our approach has the potential to greatly improve the field of non-volatile memory.”

Detecting tampering with radio waves

Researchers from Ruhr-Universität Bochum (RUB), the Max Planck Institute for Security and Privacy, and Physec propose a way to detect hardware tampering using radio waves.

The team’s method can monitor an entire system. It uses two antennas, a transmitter and receiver. The transmitter sends out a special radio signal that spreads everywhere in the system and is reflected by the walls and computer components. These reflections cause a signal to reach the receiver that is as characteristic of the system as a fingerprint.

Small changes to the system will change the fingerprint. In tests, the team equipped a conventional computer with radio antennas and punctured its housing with holes at regular intervals. Through these holes, the researchers let a fine metal needle penetrate the inside of the system and checked whether they notice the change in the radio signal. In the process, they varied the thickness of the needle, the position and the depth of penetration.

“A unique aspect of our approach is that we are carrying out the experiment while the computer is running,” said Johannes Tobisch, a PhD student at RUB.

“The fans are like little hoovers and the processor is like a heater,” said Paul Staat, a PhD student at RUB. These fluctuations in the ambient conditions affect the radio signal. The researchers measured such disturbances and factored them out in order to determine whether fluctuations in the signal are legitimate or the result of manipulation.

With the computer running, they were able to reliably detect the penetration of a needle 0.3 millimeters thick from a penetration depth of one centimeter. The system still detected a needle that was only 0.1 millimeter thick, but not in all positions. “The closer the needle is to the receiving antenna, the easier it is to detect,” said Staat.

“Therefore, in practical applications, it makes sense to think carefully about where you place the antennas,” added Tobisch. “They should be as close as possible to the components that require a high degree of protection.”

The researchers say the approach is suitable for both high-security applications and everyday protection. As well as using expensive high-precision measuring technology for recording the fingerprint, they also compared the radio signal with simple technology that sells for a handful of euros. The low-cost version also worked, with a slightly lower hit rate. “It’s always a compromise between cost and accuracy,” said Staat.

Another challenge is changes in ambient conditions, as temperature or humidity changes could affect the radio fingerprint. “We hope to tackle such problems in the future with the help of machine learning,” said Tobisch.