Resource requirements for digital computations on electrooptical systems (original) (raw)
Related papers
Computational Complexity and Efficiency in Electro-Optical Computing Systems
Ia. Z&ST UUT1OI AVALAUIETY STATIMENT 12. OISTMUTION C004 APPROVED FOR PUBLIC RELEASE: DISTRIBUTION IS UNLIMITED 13. ABSTRACT (MAuinum 200wacrl (1) To develop robust theoretical model for a wide class of electro-optical computing systems (2) To extend the known capabilities, by design of new, more efficient algorithms for electro-otpical computing using less time, volume and energy. In particular, to develop efficient algorithms that use optimal combinations of time, volume and energy on electro-optical computing systems (3) To determine the fundamental theoretical limitations and capabilities of electro-optical computing systems. In particular, to determine lower bounds on tradeoffs between volume, time, and other esources (such as energy) of any electro-optical computing system to solve fundamental problems. 14. SUBJECT T1MB I& 0F PAGES ImLC OO O 17. CU T S ASSI I ANC IS SICUFTY DASSWSCATN It. SEGJIY C IPDAT1OW A UTI OF AIED
Parallel and Sequential Optical Computing
Lecture Notes in Computer Science, 2008
We present a number of computational complexity results for an optical model of computation called the continuous space machine. We also describe an implementation for an optical computing algorithm that can be easily defined within the model. Our optical model is designed to model a wide class of optical computers, such as matrix vector multipliers and pattern recognition architectures. It is known that the model solves intractable PSPACE problems in polynomial time, and NC problems in polylogarithmic time. Both of these results use large spatial resolution (number of pixels). Here we look at what happens when we have constant spatial resolution. It turns out that we obtain similar results by exploiting other resources, such as dynamic range and amplitude resolution. However, with certain other restrictions we essentially have a sequential device. Thus we are exploring the border between parallel and sequential computation in optical computing. We describe an optical architecture for the unordered search problem of finding a one in a list of zeros. We argue that our algorithm scales well, and is relatively straightforward to implement. This problem is easily parallelisable and is from the class NC. We go on to argue that the optical computing community should focus their attention on problems within P (and especially NC), rather than developing systems for tackling intractable problems.
We consider optical computers that encode data using images and compute by transforming such images. We give an overview of a number of such optical computing architectures, including descriptions of the type of hardware commonly used in optical computing, as well as some of the computational efficiencies of optical devices. We go on to discuss optical computing from the point of view of computational complexity theory, with the aim of putting some old, and some very recent, results in context. Finally, we focus on a particular optical model of computation called the continuous space machine. We describe some results for this model including characterisations in terms of well-known complexity classes.
Parallel Algorithms for Image Processing on OMC
IEEE Transactions on Computers, 1991
In this paper, we study a class of VLSI organizations with optical interconnects for fast solutions to several image processing tasks. The organization and operation of these architectures are based on a generic model called OMC, which is proposed to understand the computational limits in using free space optics in VLSI parallel processing systems. The relationships between OMC and shared memory models are discussed in this paper. Also, three physical implementations of OMC are presented. Using OMC, we present several parallel algorithms for fine grain image computing. We categorize our results in the following order. First, we present a set of processor efficient optimal O(log N) algorithms and a set of constant time algorithms for finding geometric properties of digitized images. Finally, we focus on special purpose designs tailored to meet both the computation and communication needs of problems such as those involving irregular sparse matrices.
Massively parallel processing using optical interconnects: introduction to the feature issue
Applied Optics, 1998
This chapter discusses why and how optical communication technology might be used in future large scale parallel processing systems. Commercial systems currently being built can interconnect thousands of processors, and future systems may be even larger. Systems with thousands of processors are often called massively parallel systems. In these systems, the ability to e ciently communicate and share data between processors is critical to obtaining high performance. While optical technologies related to computation have been under investigation for many years, the technologies closest to practical implementations are those that can be applied to transmitting and processing optical signals. The use of optical technology to provide communication links between a large number of electronic processors holds promise for constructing very high performance parallel computer systems in the future.
Space-efficient optical computing with an integrated chip diffractive neural network
Nature Communications, 2022
Large-scale, highly integrated and low-power-consuming hardware is becoming progressively more important for realizing optical neural networks (ONNs) capable of advanced optical computing. Traditional experimental implementations need N2 units such as Mach-Zehnder interferometers (MZIs) for an input dimension N to realize typical computing operations (convolutions and matrix multiplication), resulting in limited scalability and consuming excessive power. Here, we propose the integrated diffractive optical network for implementing parallel Fourier transforms, convolution operations and application-specific optical computing using two ultracompact diffractive cells (Fourier transform operation) and only N MZIs. The footprint and energy consumption scales linearly with the input data dimension, instead of the quadratic scaling in the traditional ONN framework. A ~10-fold reduction in both footprint and energy consumption, as well as equal high accuracy with previous MZI-based ONNs was ...
Digital cnn with optical and electronic processing
1999
Abstract We present the preliminary results of an optoelectronic CNN system composed of two parts: a morphological nonlinear processor, which implements optical interconnections and filters, and a digital cellular processor with programmable electronic interconnections. The morphological processor has the form of a telecentric correlator with GaAs transceiver arrays composed of 8x8 active elements arranged in differential pairs.
Implications of interconnection theory for optical digital computing
Applied Optics, 1992
Heat removal, rather than finite interconnect density, is the major mechanism that limits how densely we can pack three-dimensional computing systems of increasing numbers of elements. Thus highly interconnected approaches can be employed without a further increase in system size. The use of optical interconnections for implementing the longer connections of such systems is advantageous. In fact, if the optical communication energy is sufficiently low and large-bit repetition rates are employed, conductors are useful for only the shortest connections and can be dispensed with altogether with little disadvantage. This justifies consideration of an optical digital computer. This paper is an initial attempt to understand whether and when an all-optical digital computer may prove useful. Several researchers have addressed this issue in the past, often with negative conclusions. We believe that an increasing understanding regarding the importance of communication in computing and the realization that the architectural-logical construction of a computing system can no longer be divorced from its physical construction justifies a reevaluation of previous arguments and a search for hitherto unexplored perspectives.
Parallel Algorithms For Image Convolution
In this paper, we compare the Redundant Boundary Computation (RBC) algorithm for convolution with traditional parallel methods. This algorithm dramatically reduces the communication cost for the the computation in certain environments. We theoretically and experimentally study the conventional parallel algorithm and the RBC algorithm. First, we discuss the performance of these parallel algorithms, focusing on the execution time, efficiency, speedup and scalability. Then we compare their performance to each other. Finally, we present experimental results and analysis. Our objective is to develop methods to select the correct algorithm depending on machine and problem parameters. Keywords: analysis of algorithms, parallel algorithms, image convolution, image processing, parallel performance. 1 Introduction Increasing demands for computing power have led to rapid improvements in parallel processing hardware [5, 14]. Several techniques have been used to increase the speed of parallel co...