Rapid SoC Design: On Architectures, Methodologies and Frameworks

Abstract

Modern applications like machine learning, autonomous vehicles, and 5G networking require an order of magnitude boost in processing capability. For several decades, chip designers have relied on Moore’s Law - the doubling of transistor count every two years to deliver improved performance, higher energy efficiency, and an increase in transistor density. With the end of Dennard’s scaling and a slowdown in Moore’s Law, system architects have developed several techniques to deliver on the traditional performance and power improvements we have come to expect. More recently, chip designers have turned towards heterogeneous systems comprised of more specialized processing units to buttress the traditional processing units. These specialized units improve the overall performance, power, and area (PPA) metrics across a wide variety of workloads and applications. While the GPU serves as a classical example, accelerators for machine learning, approximate computing, graph processing, and database applications have become commonplace. This has led to an exponential growth in the variety (and count) of these compute units found in modern embedded and high-performance computing platforms. The various techniques adopted to combat the slowing of Moore’s Law directly translates to an increase in complexity for modern system-on-chips (SoCs). This increase in complexity in turn leads to an increase in design effort and validation time for hardware and the accompanying software stacks. This is further aggravated by fabrication challenges (photo-lithography, tooling, and yield) faced at advanced technology nodes (below 28nm). The inherent complexity in modern SoCs translates into increased costs and time-to-market delays. This holds true across the spectrum, from mobile/handheld processors to high-performance data-center appliances. This dissertation presents several techniques to address the challenges of rapidly birthing complex SoCs. The first part of this dissertation focuses on foundations and architectures that aid in rapid SoC design. It presents a variety of architectural techniques that were developed and leveraged to rapidly construct complex SoCs at advanced process nodes. The next part of the dissertation focuses on the gap between a completed design model (in RTL form) and its physical manifestation (a GDS file that will be sent to the foundry for fabrication). It presents methodologies and a workflow for rapidly walking a design through to completion at arbitrary technology nodes. It also presents progress on creating tools and a flow that is entirely dependent on open-source tools. The last part presents a framework that not only speeds up the integration of a hardware accelerator into an SoC ecosystem, but emphasizes software adoption and usability.