Research on Massively Parallel Autonomous Multi-Processor System and Neural Network

在这项研究中，我们提出了一种用于计算机视觉和神经网络的大规模并行自治多处理器（AMP）。该系统具有大约10^5个处理器元件（PE），这些处理器元件通过通信网络彼此连接。每个PE基于数据流控制机制异步运行。在数据流控制机制中，根据数据依赖性，所有操作都并行执行，无论它们是否在同一处理器中执行。这简化了以大规模并行方式实现神经网络和计算机视觉模型。在这项研究中，我们使用软件模拟器在寄存器传输级模拟系统，估计了AMP系统的性能。这表明，该系统提供了一个高的计算性能的基础上的高并行性，不仅为低级别的图像处理，但神经网络和中间级别的图像处理。我们还提出了一种基于函数式编程语言的编程语言，它提高了大规模应用程序的软件生产率。

In this research, we have proposed a massively parallel Autonomous Multi-Processor (AMP) for computer vision and neural networks. the system has about 10^5 processor elements (PE's) connected to one another via communication network. Each PE runs asynchronously based on a data flow control mechanism. In the data flow control mechanism, all operations are executed in parallel, according to data dependency, whether they are executed in the same processor or not. This simplifies to implement neural networks and computer vision models in a massively parallel way.In this research, we have estimated the performance of the AMP system using a software simulator which simulates the system at the register transfer level. This shows that the system provides a high computation performance based on the high parallelism for not only low level image processing but for neural networks and intermediate level image processing. Also we have proposed a programming language based on functional programming language, which improves the software productivity for large scale applications.