UC Berkeley

As Moore’s Law slows and process scaling yields only small returns, computer architecture and design are poised to undergo a renaissance. This thesis brings the productivity of modern software tools to bear on the design of future energy-efficient hardware architectures.

In particular, it targets one of the most difficult design tasks in the hardware domain: Coherent hierarchies of on-chip caches. I have extended the capabilities of Chisel, a new hardware description language, by providing libraries for hardware developers to use to describe the configuration and behavior of such memory hierarchies, with a focus on the cache coherence protocols that work behind the scenes to preserve their abstraction of global shared memory. I discuss how the methods I provide enable productive and extensible memory hierarchy design by separating the concerns of different hierarchy components, and I explain how this forms the basis for a generative approach to agile hardware design.

This thesis describes a general framework for context-dependent parameterization of any hardware generator, defines a specific set of Chisel libraries for generating extensible cache-coherent memory hierarchies, and provides a methodology for decomposing high-level descriptions of cache coherence protocols into controller-localized, object-oriented transactions.

This methodology has been used to generate the memory hierarchies of a lineage of RISC-V chips fabricated as part of the ASPIRE Lab’s investigations into application-specific processor design.