Domain-Specific Acceleration: From Efficient Vision Processing Hardware to High-Performance Quantum Computing Software

PASSCODE: 000000

With the end of Dennard scaling and the decline of Moore’s law, performance and efficiency gains through pure semiconductor technology advancements have diminished. One promising path that remains for significant improvements is domain-specific acceleration (DSA), which involves designing optimized software and hardware tailored to specific application domains. In an effort to extend the boundaries of DSA for less extensively studied domains, this dissertation explores DSA designs for three application areas: Image compression, robotic vision, and quantum circuit simulation (QCS).

Firstly, an ultra-low-power H.264/AVC intra-frame image compression accelerator is developed for IoT imaging systems, optimizing energy consumption and latency. Secondly, RoboVisio, a micro-robot vision domain-specific system-on-chip (SoC) for autonomous navigation, is introduced. It features a novel hybrid processing element for efficient processing of both classic vision tasks and convolutional neural networks, achieving significant energy efficiency improvements over state-of-the-art machine learning SoCs based on non-volatile memory. Lastly, to facilitate future QCS hardware accelerator designs, the dissertation presents Fast Tensor Decision Diagram (FTDD), a novel open-source software framework for QCS. It leverages Tensor Decision Diagrams (TDD) to eliminate redundancy and achieve substantial speedups. FTDD also introduces innovative algorithms and data structures for efficient TDD operations.

CHAIR: Professor Dennis Sylvester