UC Berkeley

Application-specific integrated circuit (ASIC) signal processors are necessary to achieve the high performance and low power requirements of modern applications, but long development times are one hurdle contributing to their declining adoption. A significant percentage of their development time goes into the design and verification of the architecture, with the remainder consumed by back-end ASIC flow work and chip testing. Agile hardware principles, borrowed from a similar successful software approach and previously applied to general-purpose processors, offer a promising solution to continuing the development of signal processing systems on a chip (SoCs).

This work presents a digital signal processing SoC design framework that, when coupled with agile design principles, supports rapid prototyping and designing of ASICs for signal processing applications. First, applications and existing ASIC solutions are explored and analyzed in Chapter 2 to glean useful properties and trends. From this, Chapter 3 proposes a model for a generic signal processing SoC is developed. Next, a new Chisel generator design framework is presented in Chapter 4. Chisel is a hardware construction language written as a DSL in Scala, allowing for high-level and functional programming use when designing hardware. This framework couples a general-purpose processor with a signal processing accelerator, and much of the library code for connectivity, memory mapping, and programming is made available. This framework, when coupled with an agile design process, supports rapid development of ASICs. The accelerator performs streaming signal processing to offload high-throughput computational kernels from the CPU. As processing elements for the desired application are produced, processing moves from the CPU to the accelerator. Low-rate processing tasks are computed on the CPU, meaning tape-out occurs on time and produces a working chip able to perform the entire application.

The methodology and proposed agile design process were validated on two separate chips in Chapters 5 and 6, spanning two applications and two process nodes. The ASIC spectrometer (Splash2), for which the RTL was designed in eight weeks by one person, demonstrates the power of Chisel to rapidly construct processing element generators. These generators were then improved and the parameters adjusted as physical design and timeline constraints imposed new restrictions. The radar receive processor design fleshed out the generator framework details. A significantly larger design, this chip required about 300 engineering-weeks of work over 9 months, equivalent to a team of 8 engineers working full time. About 30% of that time was spent designing the framework and reusable processing elements. This represents a 56% reduction in development time compared to the estimated 14.4 months from standard practices (excluding time for framework design, fabrication, and testing). Both efforts produced working chips competitive against state-of-the-art custom ASICs in terms of performance, power, and capabilities.