The main characteristic of Reconfigurable Computing is the presence of hardware that can be reconfigured to implement specific functionality more suitable for specially tailored hardware than on a simple uniprocessor. Reconfigurable computing systems join microprocessors and programmable hardware in order to take advantage of the combined strengths of hardware and software and have been used in applications ranging from embedded systems to high performance computing. Many of the fundamental theories have been identified and used by the Hardware/Software Co-Design research field. Although the same background ideas are shared in both areas, they have different goals and use different approaches.This book is intended as an introduction to the entire range of issues important to reconfigurable computing, using FPGAs as the context, or "computing vehicles" to implement this powerful technology. It will take a reader with a background in the basics of digital design and software programming and provide them with the knowledge needed to be an effective designer or researcher in this rapidly evolving field.

· Treatment of FPGAs as computing vehicles rather than glue-logic or ASIC substitutes
· Views of FPGA programming beyond Verilog/VHDL
· Broad set of case studies demonstrating how to use FPGAs in novel and efficient ways

Reconfigurable Computing: The Theory and Practice of FPGA-Based Computation

Steady advances in VLSI technology and design tools have extensively expanded the application domain of digital signal processing over the past decade. While application-specific integrated circuits (ASICs) and programmable digital signal processors (PDSPs) remain the implementation mechanisms of choice for many DSP applications, increasingly new system implementations based on reconfigurable computing are being considered. These flexible platforms, which offer the functional efficiency of hardware and the programmability of software, are quickly maturing as the logic capacity of programmable devices follows Moore's Law and advanced automated design techniques become available. As initial reconfigurable technologies have emerged, new academic and commercial efforts have been initiated to support power optimization, cost reduction, and enhanced run-time performance.

This paper presents a survey of academic research and commercial development in reconfigurable computing for DSP systems over the past fifteen years. This work is placed in the context of other available DSP implementation media including ASICs and PDSPs to fully document the range of design choices available to system engineers. It is shown that while contemporary reconfigurable computing can be applied to a variety of DSP applications including video, audio, speech, and control, much work remains to realize its full potential. While individual implementations of PDSP, ASIC, and reconfigurable resources each offer distinct advantages, it is likely that integrated combinations of these technologies will provide more complete solutions.

http://www.umiacs.umd.edu/~joseph/classes/enee640/fpga-dsp-survey.pdf

Reconfigurable Computing for Digital Signal Processing: A Survey

Moore's Law states that the number of transistors on a device doubles every two years; however, it is often (mis)quoted based on its impact on CPU performance. This important corollary of Moore's Law states that improved clock frequency plus improved architecture yields a doubling of CPU performance every 18 months. This paper examines the impact of Moore's Law on the peak floating-point performance of FPGAs. Performance trends for individual operations are analyzed as well as the performance trend of a common instruction mix (multiply accumulate). The important result is that peak FPGA floating-point performance is growing significantly faster than peak floating-point performance for a CPU.

/pdf/fpgas-vs-cpus-trends-in-peak-floating-point-performance-w1q6p0ta0k.pdf

FPGAs vs. CPUs: trends in peak floating-point performance

The development of neural networks has now reached the stage where they are employed in a large variety of practical contexts. However, to date the majority of such implementations have been in software. While it is generally recognised that hardware implementations could, through performance advantages, greatly increase the use of neural networks, to date the relatively high cost of developing Application-Specific Integrated Circuits (ASICs) has meant that only a small number of hardware neurocomputers has gone beyond the research-prototype stage. The situation has now changed dramatically: with the appearance of large, dense, highly parallel FPGA circuits it has now become possible to envisage putting large-scale neural networks in hardware, to get high performance at low costs. This in turn makes it practical to develop hardware neural-computing devices for a wide range of applications, ranging from embedded devices in high-volume/low-cost consumer electronics to large-scale stand-alone neurocomputers. Not surprisingly, therefore, research in the area has recently rapidly increased, and even sharper growth can be expected in the next decade or so. Nevertheless, the many opportunities offered by FPGAs also come with many challenges, since most of the existing body of knowledge is based on ASICs (which are not as constrained as FPGAs). These challenges range from the choice of data representation, to the implementation of specialized functions, through to the realization of massively parallel neural networks; and accompanying these are important secondary issues, such as development tools and technology transfer. All these issues are currently being investigated by a large number of researchers, who start from different bases and proceed by different methods, in such a way that there is no systematic core knowledge to start from, evaluate alternatives, validate claims, and so forth. FPGA Implementations of Neural Networks aims to be a timely one that fill this gap in three ways: First, it will contain appropriate foundational material and therefore be appropriate for advanced students or researchers new to the field. Second, it will capture the state of the art, in both depth and breadth and therefore be useful researchers currently active in the field. Third, it will cover directions for future research, i.e. embryonic areas as well as more speculative ones.

FPGA Implementations of Neural Networks

We introduce a 64-bit ANSI/IEEE Std 754-1985 floating point design of a hardware matrix multiplier optimized for FPGA implementations. A general block matrix multiplication algorithm, applicable for an arbitrary matrix size is proposed. The algorithm potentially enables optimum performance by exploiting the data locality and reusability incurred by the general matrix multiplication scheme and considering the limitations of the I/O bandwidth and the local storage volume. We implement a scalable linear array of processing elements (PE) supporting the proposed algorithm in the Xilinx Virtex II Pro technology. Synthesis results confirm a superior performance-area ratio compared to related recent works. Assuming the same FPGA chip, the same amount of local memory, and the same I/O bandwidth, our design outperforms related proposals by at least 1.7X and up to 18X consuming the least reconfigurable resources. A total of 39 PEs can be integrated into the xc2vp125-7 FPGA, reaching performance of, e.g., 15.6 GFLOPS with 1600 KB local memory and 400 MB/s external memory bandwidth.

/pdf/64-bit-floating-point-fpga-matrix-multiplication-3hha35fr6i.pdf

64-bit floating-point FPGA matrix multiplication

The use of reconfigurable hardware to perform high precision operations such as IEEE floating point operations has been limited in the past by FPGA resources. We discuss the implementation of IEEE single precision floating-point multiplication and addition. Then, we assess the practical implications of using these operations in the Xilinx 4000 series FPGAs considering densities available now and scheduled for the near future. For each operation, we present space requirements and performance information. This is followed by a discussion of an algorithm, matrix multiplication, based on these operations, which achieves performance comparable to conventional microprocessors. Algorithm implementation options and their performance implications are discussed and corresponding measured results are given.

S. P. McMillan

Papers

A re-evaluation of the practicality of floating-point operations on FPGAs

Implementation of IEEE single-precision floating-point operations on FPGAs (abstract)