Applications of reprogrammability in algorithm acceleration

Abstract

This doctoral thesis consists of an introductory part and eight appended publications, which deal with hardware-based reprogrammability in algorithm acceleration with a specific emphasis on the possibilities offered by modern large-scale Field Programmable Gate Arrays (FPGAs) in computationally demanding applications.

The historical evolution of both the theoretical and technological paths culminating in the introduction of reprogrammable logic devices is first outlined. This is followed by defining the commonly used terms in the thesis. The reprogrammable logic market is surveyed, and the architectural structures and the technological reasonings behind them are described in detail. As reprogrammable logic lies between Application Specific Integrated Circuits (ASICs) and general-purpose microprocessors in the implementation spectrum of electronics systems, special attention has been paid to differentiate these three implementation approaches. This has been done to emphasize, that reprogrammable logic offers much more than just a low-volume replacement for ASICs.

Design systems for reprogrammable logic are investigated, as the learning curve associated with them is the main hurdle for software-oriented designers for using reprogrammable logic devices. The theoretically important topic of partial reprogrammability is described in detail, but it is concluded, that the practical problems in designing viable development platforms for partially reprogrammable systems will hinder its wide-spread adoption.

The main technical, design-oriented, and economic applicability factors of reprogrammable logic are laid out. The main advantages of reprogrammable logic are their suitability for fine-grained bit-level parallelizable computing with a short time-to-market and low upfront costs. It is also concluded, that the main opportunities for reprogrammable logic lie in the potential of high-level design systems, and the ever-growing ASIC design gap. On the other hand, most power-conscious mass-market portable products do not seem to offer major new market potential for reprogrammable logic.

The appended publications are examined and compared to contemporaneous research at other research institutions. The conclusion is that for relatively wide classes of well-defined computation problems, reprogrammable logic offers a more efficient solution than a software-centered approach, with a much shorter production cycle than is the case with ASICs.