Current Research Projects
Principal Investigators: David Peak (Physics) and Aravind Dasu (ECE)
A marked difference between technological and biological systems is how they cope with irregularity. Specialized equipment or procedures are frequently required to detect and correct errors in technological systems, whereas biological systems often manage to function quite successfully in highly variable environments without any obvious error correction capability. The research proposed here specifically focuses on how this latter trick might be mimicked in "self-healing" computational networks - ones in which error-detection and -correction are accomplished without central control - and how such networks might be usefully employed to mitigate both single and multi event upsets (SEU and MEU) caused by transient and permanent classes of errors that are increasingly encountered as IC technologies incorporate shrinking transistor sizes and complex 3D routing in stacks. To ground this research in reality, we propose to implement such self-healing techniques for image processing Cellular Neural Networks (CNNs) on state of the art Xilinx FPGAs and demonstrate their self-recovery from a variety of faults. An ancillary goal will be to optimize the design for extremely low area and testability of almost every bit in the configuration bit stream for controlled corruption and observation of system emergence from those errors.
COUPLED THERMAL-ELECTRIC-STRUCTURAL ANALYSES OF INTERCONNECTS ON 3D-ICs - A Study of Fundamental Physics of Interconnect Failure
Principal Investigators: Leijun Li (MAE)
We propose to study the design of interconnects for 3D ICs using a thermal-mechanical analysis approach. The temperature and stress fields of 3D ICs will be simulated using a transient, coupled- field, finite element method. Various interconnect and solid-metal thermal spreader layouts will be compared. The effect of defects and failure in the vertical via chain structures on thermal management will be quantitatively analyzed. The results will be directly applicable to the IC industry for optimizing the interconnect design and fabrication.
Principal Investigators: Leila Ladani (MAE)
Virtual qualification (VQ) approach is proposed to be used in detecting the failure mechanism of new technology of 3-D IC packages for the most common type of loading which is power and thermal cycling. This approach will use the Physics of Failure (PoF) principles to determine the 3-D temperature distribution, thermal stress, failure mechanisms, and fatigue life of the interconnects and active layers of the package under the applied loading conditions. The outcomes of the proposed plan are design recommendations and enhancements that will help reducing the stress levels in Copper interconnects and different layers of the package and increases the durability and reliability of the package. This research will also provide an estimation of reliability for copper interconnects and active layers of package and provide a database of active physical mechanisms that will cause failure during the life cycle of these devices. Furthermore, it will provide the manufacturer with a qualification plan that will be applicable for different type of devices made by this type of technology.
SOFT REED-SOLOMON DECODING
Principal Investigators: Todd Moon (ECE)
Reed-Solomon codes have been widely employed in memory systems, providing the ability to correct errors by the controlled introduction of redundancy. As physical dimensions of memory cells decrease, there is a possibility of increasing error rates. To compensate for this, improved error correction decoding performance is sought. In this research project, new decoding algorithms are being developed which exploit soft information for Reed-Solomon decoding. Unlike other “soft” RS decoding algorithms, which only approximately map received data to probabilities, the approaches here explicitly employ probabilities computed directly from received data. The first approach develops a set of “soft key equations” from the encoding operation. These soft key equations have a sparse representation, which allows for their solution using belief propagation. The second approach employs a “decoding by encoding” strategy, in which an approximate soft ML solution is sought.