Whatmough, PN; (2012) Timing-Error Tolerance Techniques for Low-Power DSP: Filters and Transforms. Doctoral thesis , UCL (University College London).

Preview

PDF Paul Whatmough EngD Thesis_Aug2012.pdf Available under License : See the attached licence file. Download (3MB)

Abstract

Low-power Digital Signal Processing (DSP) circuits are critical to commercial System-on-Chip design for battery powered devices. Dynamic Voltage Scaling (DVS) of digital circuits can reclaim worst-case supply voltage margins for delay variation, reducing power consumption. However, removing static margins without compromising robustness is tremendously challenging, especially in an era of escalating reliability concerns due to continued process scaling. The Razor DVS scheme addresses these concerns, by ensuring robustness using explicit timing-error detection and correction circuits. Nonetheless, the design of low-complexity and low-power error correction is often challenging. In this thesis, the Razor framework is applied to fixed-precision DSP filters and transforms. The inherent error tolerance of many DSP algorithms is exploited to achieve very low-overhead error correction. Novel error correction schemes for DSP datapaths are proposed, with very low-overhead circuit realisations. Two new approximate error correction approaches are proposed. The first is based on an adapted sum-of-products form that prevents errors in intermediate results reaching the output, while the second approach forces errors to occur only in less significant bits of each result by shaping the critical path distribution. A third approach is described that achieves exact error correction using time borrowing techniques on critical paths. Unlike previously published approaches, all three proposed are suitable for high clock frequency implementations, as demonstrated with fully placed and routed FIR, FFT and DCT implementations in 90nm and 32nm CMOS. Design issues and theoretical modelling are presented for each approach, along with SPICE simulation results demonstrating power savings of 21 – 29%. Finally, the design of a baseband transmitter in 32nm CMOS for the Spectrally Efficient FDM (SEFDM) system is presented. SEFDM systems offer bandwidth savings compared to Orthogonal FDM (OFDM), at the cost of increased complexity and power consumption, which is quantified with the first VLSI architecture.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item