A software-defined radio (SDR) ideally allows all of the parameters of a radio to be programmed dynamically. While the circuits on the baseband of radios are easily tunable, the RF front-end is not.
Our research has focused on developing methods of manipulating the impedance seen by the antenna using baseband circuits. We have achieved this using passive mixers connected directly to the antenna (without an LNA) with performance competitive with traditional radios. Our group has also developed circuit architectures to enable full-duplex communications for efficient spectrum usage.
Hetero-junction Bipolar Transistor (HBT) Based Mixer-First Radio
Growing interest in overcoming the limitations of Moore’s law has led to the design and fabrication of novel devices that beat the short-channel effects inherent to smaller, faster CMOS transistors. Recently, InP HBTs have shown fT in the terahertz range, whereas state-of-the-art CMOS for software defined radios have fT in the low gigahertz, incapable of utilizing the extended frequency range of the HBT devices. Our research focuses on preserving the impedance transparency property of CMOS mixer-first radios using non-bidirectional HBTs such that we can achieve high out-of-band linearity beyond 2 GHz. We have demonstrated 10 dBm 1dB-compression of an in-band tone by out-of-band blockers at 10 GHz, and are currently working towards breaking the trade-off between frequency, power, noise, and linearity for widely-tunable microwave to mm-wave radio.
Programmable Duplexing Radios
[Hazal Yüksel, Zach Boynton, Logan Horowitz, and Michael Rivera]
Due to the proliferation of smartphones, spectrum is becoming an ever scarcer resource. There have been different attempts at alleviating the crunch, but duplexing solutions show promising results.
Our research focuses on ultra flexible, software defined, widely tunable solutions to the duplexing problem. We use a distributed amplifier based transmitter with a transmission line where we allow for adjustable gain and phase inputs of the sub-transmitters to null the TX signal at the receiver and add it constructively at the antenna port. We combine this technique with a frequency selective degeneration to selectively degenerate the gain of the amplifier in the RX band, reducing the noise the sub-transmitters contribute to the RX noise figure.
Angle Sensitive Pixels
Our original goal was to detect and localize luminescent or fluorescent sources in a 3D matrix (such as tissue) or solution without using a lens and at low cost and size. The key idea that made this possible is the Angle-Sensitive Pixel (ASP). This structure, which can be fabricated entirely in a standard CMOS process, detects the angle of incident light by capturing the diffraction pattern formed as the light passes through a pair of stacked diffraction gratings made from metal in the routing layers of the CMOS process. An array of ASPs can capture the incident 4D light field, allowing us to digitally refocus images post-capture, estimate feature depth within the visual scene, and even localize multiple light sources in a 3D volume. Since then, we have researched ASP functionality at multiple levels of abstraction, from opto-electronic device development inside of standard CMOS, to circuit and system design to preprocess data from large numbers of ASPs, and further to machine-learning-based signal processing for plenoptic imaging. ASPs can extract information from the scene directly that would otherwise need to be computed post-capture, at the cost of power and system bandwidth. We have used this sensitivity to optically compute the first layer of convolutional neural networks for energy-efficient deep learning, and are now experimenting with ASPs as ultra-low-power sensors for power-starved robotic systems and sensor networks.
Near-Zero Power (N-Zero)
[Melissa White and Robin Ying]
Our signal processing research positions us at the intersection of machine learning and low-power hardware design. Low-power chips can “listen” passively to their surroundings for extended periods, and only “wake up” at a particular trigger, having diverse applications in defense, environmental monitoring, and health.
The problems of optimization or classification are usually the purview of computer scientists who have large amounts of computing power, and design of low-power circuits focuses on other classes of problems. There is an emerging design space at the intersection of these two – where hardware power consumption places limits on allowable operations, and hardware design is guided by algorithmic approaches. We aim to address these simultaneously, using a variety of tools from classical machine learning, data processing, and programmable low-power CMOS design.
Micro-scale Opto-electrically Transduced Electrode (MOTE)
[Sunwoo Lee, Aasta P. Gandhi, Paige Trexel,
Frances Koback, and Emily P. Kuck]
Recording neural activity in live animals in vivo poses several challenges. Electrical techniques typically require electrodes to be tethered to the outside world directly via a wire, or indirectly via an RF Coil, which is much larger than the electrodes themselves. Tethered implants result in residual motion between neurons and electrodes as the brain moves, and limits our ability to measure from peripheral nerves in moving animals, especially in smaller organisms such as zebra fish or fruit flies. On the other hand, optical techniques, which are becoming increasingly powerful, are nonetheless often limited to subsets of neurons in any given organism, impeded by scattering of the excitation light and emitted fluorescence, and limited to low temporal resolution. MOTE entails the electronics for an untethered electrode unit, powered by, and communicating through a microscale optical interface, combining many benefits of optical techniques with high temporal-resolution recording of electrical signals.
By combining very compact amplification and digitization of neural signals with angle sensitive pixels, we are building multi-hundred active electrode arrays intermingled with lensless imaging for simultaneous opto-electronic recording of neural tissue. We are also building on this, combined with techniques we have developed for low power, low noise neural recording to ultimately develop implantable opto-electronic neural probes.