Variation-aware high-level DSP circuit design optimisation framework for FPGAs
File(s)
Author(s)
Policarpo Duarte, Rui
Type
Thesis
Abstract
The constant technology shrinking and the increasing demand for systems that operate under different power profiles with the maximum performance, have motivated the work in this thesis. Modern design tools that target FPGA devices take a conservative approach in the estimation of the maximum performance that can be achieved by a design when it is placed on a device, accounting for any variability in the fabrication process of the device.
The work presented here takes a new view on the performance improvement of DSP designs by pushing them into the error-prone regime, as defined by the synthesis tools, and by investigating methodologies that reduce the impact of timing errors at the output of the system.
In this work two novel error reduction techniques are proposed to address this problem. One is based on reduced-precision redundancy and the other on an error optimisation framework that uses information from a prior characterisation of the device. The first one is a generic architecture that is appended to existing arithmetic operators. The second defines the high-level parameters of the algorithm without using extra resources. Both of these methods allow to achieve graceful degradation whilst variation increases.
A comparison of the new methods is laid against the existing methodologies, and conclusions drawn on the tradeoffs between their cost, in terms of resources and errors, and their benefits in terms of throughput.
In some cases it is possible to double the performance of the design while still producing valid results.
The work presented here takes a new view on the performance improvement of DSP designs by pushing them into the error-prone regime, as defined by the synthesis tools, and by investigating methodologies that reduce the impact of timing errors at the output of the system.
In this work two novel error reduction techniques are proposed to address this problem. One is based on reduced-precision redundancy and the other on an error optimisation framework that uses information from a prior characterisation of the device. The first one is a generic architecture that is appended to existing arithmetic operators. The second defines the high-level parameters of the algorithm without using extra resources. Both of these methods allow to achieve graceful degradation whilst variation increases.
A comparison of the new methods is laid against the existing methodologies, and conclusions drawn on the tradeoffs between their cost, in terms of resources and errors, and their benefits in terms of throughput.
In some cases it is possible to double the performance of the design while still producing valid results.
Version
Open Access
Date Issued
2014-05
Date Awarded
2014-08
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Bouganis, Christos-Savvas
Sponsor
Fundacao para a Ciencia e a Tecnologia
Grant Number
SFRH/BD/69587
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)