Heterogeneous Chip Multiprocessor: Data Representation, Mixed-Signal Processing Tiles, and System Design

With the emergence of big data, the need for more computationally intensive processors that can handle the increased processing demand has risen. Conventional computing paradigms based on the Von Neumann model that separates computational and memory structures have become outdated and less efficient for this increased demand. As the speed and memory density of processors have increased significantly over the years, these models of computing, which rely on a constant stream of data between the processor and memory, see less gains due to finite bandwidth and latency. Moreover, in the presence of extreme scaling, these conventional systems, implemented in submicron integrated circuits, have become even more susceptible to process variability, static leakage current, and more. In this work, alternative paradigms, predicated on distributive processing with robust data representation and mixed-signal processing tiles, are explored for constructing more efficient and scalable computing systems in application specific integrated circuits (ASICs). The focus of this dissertation work has been on heterogeneous chip multi-processor (CMP) design and optimization across different levels of abstraction. On the level of data representation, a different modality of representation based on random pulse density modulation (RPDM) coding is explored for more efficient processing using stochastic computation. On the level of circuit description, mixed-signal integrated circuits that exploit charge-based computing for energy efficient fixed point arithmetic are designed. Consequently, 8 different chips that test and showcase these circuits were fabricated in submicron CMOS processes. Finally, on the architectural level of description, a compact instruction-set processor and controller that facilitates distributive computing on System-On-Chips (SoCs) is designed. In addition to this, a robust bufferless network architecture is designed with a network simulator, and I/O cells are designed for SoCs. The culmination of this thesis work has led to the design and fabrication of a heterogeneous chip multi- processor prototype comprised of over 12,000 VVM cores, warp/dewarp processors, cache, and additional processors, which can be applied towards energy efficient large-scale data processing