The overall research goal in ICSRL is to comprehend the mathematical and physical models that govern nature and our interactions with it. Engineering, has always been the application of man’s understanding of natural sciences to design systems that enable a better tomorrow. In our lab, we endeavor to explore such models of information processing and computation, as well as to design hardware solutions to physically realize them.
Energy-efficient Hardware for Autonomous Systems, Machine Learning and Smart Sensing
We explore mixed-signal hardware that can enable the next generation of autonomy in intelligent systems. We have pioneered computational data-converters that can perform in-situ classification on voltage as well as time-based ADCs. Our current work encompasses hardware solutions for voltage and process scalability, memory-centric processing and low-cost interconnect technologies for supervised, unsupervised and reinforcement based learning. We collaborate with device and technology researchers to understand how post-CMOS devices can enable next-generation of autonomous and intelligent systems. We have demonstrated world’s first hardware for reinforcement learning at sub-1mW targeted for autonomous robots. Our work on “always-on” smart cameras has been widely covered by media outlets including Wired, Engadget and TechCrunch.
Key Publications: [ISSCC 2018][Symp. of VLSI Technology 2017][ASSCC 2017][BioCAS 2017][TCAS-I 2017][TCAS-II 2016][IEDM 2014]
Design of Voltage Regulators, Adaptive Clocking, and Power Management
This project explores the design of on-die power management circuits with novel voltage-regulator architectures that are suited for digital logic. The focus of the work is to develop control architectures and corresponding hardware implementations that provide stability and performance across wide dynamic ranges of operation. Our principle interests are in linear and switched capacitor regulator topologies that are compact, low overhead and highly efficient.
Key Publications: [JSSC 2018][ESSCIRC 2016][ESSCIRC 2016][ISSCC 2015][JSSC 2014][Symp. on VLSI Circuits 2012]
Computation with Non-linear Dynamical Systems
While Boolean, Von-Neumann machines have fueled the technology revolution over the last three decades, it is well recognized that brain-inspired computational models are well suited for tasks such as data classification and recognition. Our current work is at the interface of non-linear dynamics exhibited by correlated electron devices and the computational models that are made possible by such complex systems. We are engaged in active research in the “information processing” capabilities of non-linear dynamical systems where the dynamics of synchronous oscillators are stimulated, controlled and observed; and can be shown to perform tasks such as associative matching. Our work on non-linear dynamical systems that connect dynamics with algebraic graph theory has been widely covered by media outlets including Phys.Org, Engadget and TechCrunch.
Key Publications: [Nature SR 2017][Journal of Applied Physics 2015][IEDM 2014][Nature 2017] [JETCAS 2014][DAC 2014]
Energy vs Accuracy: Computational Models and Hardware for Sensors, Sensor Interfaces and Signal Processing
Designing CMOS based sensors, sensor interfaces and low-power processing of sensor data is of interest not only for environmental sensing but also for advanced human-machine interfaces. We are interested in the design and applications of on-chip sensors like temperature, voltage, and current sensors as well as in building platform sensors (audio and video) that will enable seamless interaction of human and machines. Currently, we are exploring novel hardware concepts in compressive sensing, random sampling and processing in reduced dimensions with a focus towards approximate computation. This includes understanding and exploring mathematical models as well as physical realization of such models in digital and mixed signal domains.
Key Publications: [ISSCC 2009] [JSSC 2009] [CICC 2012]
Spintronics: Beyond Charge Based Computation and Storage
Prof. Raychowdhury is an expert on device modeling, compact modeling, circuit design, and test architectures for Spin-Transfer-Torque based Magnetic RAMs (STT-MRAMs). Our research group is exploring improved device models for fast evaluation of process parameter variation in STT-MRAM with a focus towards different failure models. Current work includes identifying circuit architectures and test generation methods for designing and characterizing robust STT-RAM arrays in embedded designs. More recently, we have been exploring the application of Spintronic devices in analog and mixed signal computation.
Key Publications: [IEDM 2009] [IEDM 2010] [IEDM 2011] [TED 2011]
Design of Low-Power Digital and Mixed-Signal Circuits with Emphasis on Adaptability and Resiliency
We are interested in exploiting the limits of low energy design with adaptation and resiliency against dynamic variations and errors. The aim is to minimize design guard-bands to ensure error free operation with a need for error detection and correction in case of rare failure events. We study circuits and architecture for both logic and embedded memory.
Key Publications: [JSSC 2011] [ISSCC 2011] [Symp. on VLSI Circuits 2010]