| Title | : | Design Techniques for Low Power Tunable Approximate Circuits |
| Speaker | : | Neel Talakshi Gala (IITM) |
| Details | : | Tue, 16 Feb, 2016 3:00 PM @ BSB 361 |
| Abstract: | : | With continuously increasing transistor density, as governed by Moores law, energy efficiency is of paramount importance in designing digital systems today. As next generation systems become more mobile and embedded, the computational tasks have become more complex and sophisticated including a wide variety of applications involving multimedia processing, DSP, data mining, search and analytics, etc. These emerging applications share a key property - their computations can be executed approximately without significantly affecting the output quality. Thus, hardware designs catering to these applications can be designed with less strict constraints and guard bands allowing certain errors to propagate at the output while gaining benefits in performance and/or energy. Various approximation techniques have been proposed in literature, spanning the entire stack - from hardware to software - exploring different aspects of energy- quality/energy-performance trade-offs. However, the most critical aspect of approximating a design lies in identifying parts of the design or application which can be approximated. In this work, we propose a framework which employs an automated application-independent flow to identify which regions/parts of the design, at the gate level, can be leveraged for approximation and significant power savings. Post this analysis, we employ a series of hardware based optimizations such as voltage scaling, power- gating, Multi-Vt swap, etc. to create a tunable circuit, capable of operating in both accurate and approximate modes based on environmental conditions and user inputs. With this work we have observed upto 35% reduction in power for applications targeting videos and image processing. |
