University of British Columbia
Fast design space exploration for low-power configurable processors
Abstract
dc:descriptionCustomizable and extensible processors (commonly known as “configurable processors” or ASIPs) can provide the flexibility of off-the-shelf processors with a performance closer to that of custom logic. Manual configuration of an ASIP requires highly-specialized knowledge of computer architecture and typically results in suboptimal architectures leading to poor performance and higher costs. Ideally, the ASIP flow should be entirely automated; however, optimal solutions are only guaranteed with an exhaustive search of the design space. Unfortunately, an exhaustive search is computationally prohibitive and so the research community continues to study ways to find “good” solutions within a reasonable time. This dissertation presents new methods of design space exploration and fast architecture evaluation. These methods are intended to improve the automation and usability of ASIPs. Design space exploration is conducted using a novel approach where the design space is modeled using a small sample of points. Each sample point evaluation is expensive; however, the design space model can then be used to quickly estimate all other points in the space. Non-parametric statistics are used to construct the model and, consequently, the precise nature of the design space need not be specified a priori. This approach provides a computationally-efficient alternative to existing optimization heuristics with additional benefits that provide easy discovery of architectural trends and tradeoffs. Experiments were conducted using the proposed modeling approach to configure both the branch prediction unit (BPU) and the cache hierarchy of an embedded processor. Results showed that the approach could achieve a 1 OOx speedup while providing near optimal configurations. In addition, a fast performance estimation approach is proposed for evaluating configurations of instruction-set extensions. This approach considers pipeline effects and consequently improves the quality of results over existing approaches. This improvement is achieved while maintaining constant run-time complexity.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2008
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hallschmid, Peter
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/5283