2015 Theses Doctoral
Supervised Design-Space Exploration
Low-cost Very Large Scale Integration (VLSI) electronics have revolutionized daily life and expanded the role of computation in science and engineering. Meanwhile, process-technology scaling has changed VLSI design to an exploration process that strives for the optimal balance among multiple objectives, such as power, performance, and area, i.e. multi-objective Pareto-set optimization. Besides, modern VLSI design has shifted to synthesis-centric methodologies in order to boost the design productivity, which leads to better design quality given limited time and resources. However, current decade-old synthesis-centric design methodologies suffer from: (i) long synthesis tool runtime, (ii) elusive optimal setting of many synthesis knobs, (iii) limitation to one design implementation per synthesis run, and (iv) limited capability of digesting only component-level designs as opposed to holistic system-wide synthesis. These challenges make Design Space Exploration (DSE) with synthesis tools a daunting task for both novice and experienced VLSI designers, thus stagnating the development of more powerful (i.e. more complex) computer systems.
To address these challenges, I propose Supervised Design-Space Exploration (SDSE), an abstraction layer between a designer and a synthesis tool, aiming to autonomously supervise synthesis jobs for DSE. For system-level exploration, SDSE can approximate a system Pareto set given limited information: only lightweight component characterization is required, yet the necessary component synthesis jobs are discovered on-the-fly in order to compose the system Pareto set. For component-level exploration, SDSE can approximate a component Pareto set by iteratively refining the approximation with promising knob settings, guided by synthesis-result estimation with machine-learning models. Combined, SDSE has been applied with the three major synthesis stages, namely high-level, logic, and physical synthesis, to the design of heterogeneous accelerator cores as well as high-performance processor cores. In particular, SDSE has been successfully integrated into the IBM Synthesis Tuning System, yielding 20% better circuit performance than the original system on the design of a 22nm server processor that is currently in production.
Looking ahead, SDSE can be applied to other VLSI designs beyond the accelerator and the programmable cores. Moreover, SDSE opens several research avenues for: (i) new development and deployment platforms of synthesis tools, (ii) large-scale collaborative design engineering, and (iii) new computer-aided design approaches for new classes of systems beyond VLSI chips.
- Liu_columbia_0054D_12831.pdf binary/octet-stream 2.25 MB Download File
More About This Work
- Academic Units
- Computer Science
- Thesis Advisors
- Carloni, Luca P.
- Ph.D., Columbia University
- Published Here
- July 30, 2015