EMT/MISC Nanogrid Implementation of

NSF 08-517-新兴计算模型和技术(EMT)EMT/MISC纳米网格大规模并行算法的实现PI：Dan Hammerstrom，PSU；Co-PI：Richard Granger，达特茅斯；Co-PI：Konstantin Likharev，纽约州立大学石溪分校摘要我们的目标是研究使用全新的实现技术，以新的算法和新的架构和设计技术来增强智能计算(IC)。我们将这项工作的重点放在计算机视觉中的识别问题上。我们正在研究的算法起源于神经科学，但它们是那些算法的重要抽象，目的是保留计算的本质，同时去掉许多生物学细节。第一个假设是，这些算法构成了一种很有前途的方法，可以提高IC的水平。第二个假设是，扩展到超大型网络是智能计算的必要要求，而我们正在使用的算法确实是可扩展的。第三个假设是，对于这些算法的规模化实现，CMOS永远不会以合理的成本/性能为我们提供所需的算法扩展。因此，我们需要转向密度高得多的介质。因此，我们的第四个假设是，混合CMOL(CMOL)CMOL是最有前景的实现技术。因此，本文提出的研究目标是在CMOL中实现大规模并行的、基于统计的算法，这是我们对一般问题的解决方案。我们的方法是使用具有许多不同算法阶段的实际应用程序，并研究该阶段到CMOL的映射。每个阶段的设计范围将基于我们称为虚拟化的概念。特定实现的性能/价格由虚拟化度决定，而虚拟化度又由算法及其动态行为决定。

NSF 08-517 - Emerging Models and Technologies for Computation (EMT)EMT/MISC Nanogrid Implementation ofMassively Parallel AlgorithmsPI: Dan Hammerstrom, PSU; Co-PI: Richard Granger, Dartmouth;Co-PI: Konstantin Likharev, SUNY Stony BrookAbstractOur goal is to investigate the use of radically new implementation technology to enhance Intelligent Computing (IC) with new algorithms and new architectural and design techniques. We are focusing this work on recognition problems in computer vision. The algorithms we are studying have their origin in neuroscience, but they are significant abstractions of those algorithms, with the goal of retaining the essence of the computation while dropping many of the biological details. The 1st assumption is that these algorithms constitute a promising approach to achieving improved levels of IC. The 2nd assumption is that scaling to very large networks is a necessary requirement of intelligent computing, and the algorithms we are using do scale. The 3rd assumption is that CMOS will never give us the algorithm scaling we need at a reasonable cost/performance for scaled implementations of these algorithms. Thus we need to move to a far denser medium. Our 4th assumption then is that hybrid CMOS / nanogrids (CMOL)CMOL is the most promising implementation technology on the horizon.Consequently the goal of the research proposed here is to implement massively parallel, statistically based algorithms in CMOL, which is our solution to the general problem. Our approach is to take a real application with a number of different algorithmic stages and study the mapping of that stage to CMOL. The design spectrum for each stage will be based on a concept we call virtualization. The performance/price of a particular implementation is determined by the degree of virtualization, which in turn is determined by the algorithm and its dynamic behavior.