SHF: Medium: Collaborative Research: Photonic Neural Network Accelerator for Energy-efficient Heterogeneous Multicore Architectures

卷积神经网络和递归神经网络等深度学习结构在人工智能的许多现代应用中取得了前所未有的、有时甚至超乎人类的精度，例如图像分类和语音识别。然而，在这些耗能的机器学习体系结构中，功耗是一个主要问题，降低功耗需要设计提供更节能的硬件和机器学习算法的组合。越来越重视利用并行性和专门化来提高性能和能效。为了显著降低功耗，与电气实现相比，硅光子学被提出来提高每瓦特性能。该项目利用光子技术和异质多核来设计深度神经网络加速器，以提高各种机器学习应用中的并行性、并发性、能效和可扩展性。该项目的第一个任务是表征和识别能够实现加速器功能的光子器件，例如乘法和累加、求和和其他算术运算。然后，将表征的器件插入单层和多层光子拓扑中，以实现加速器功能。该项目的第二个任务是在提出的光子神经网络加速器上实现各种类型的深度学习架构，以最大限度地发挥光子技术提供的收益。该项目的第三个任务是构建一个广泛的模拟和建模基础设施，该基础设施结合了前两个步骤中开发的光子技术、网络体系结构、加速器功能和机器学习算法，以验证光子神经网络加速器能够显著降低能耗。因此，由于其横切性质，预计它将对下一代多核体系结构的设计产生深远影响。它将在几个领域培育新的研究方向，横跨计算机体系结构、光学技术、机器学习算法和应用。这项研究还将通过将发现与教学和培训相结合，在教育中发挥重要作用。所有的研究成果和模拟工具包将通过会议和期刊出版物以及一个专门的网站向社区传播。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Deep learning architectures such as convolutional neural networks and recurrent neural networks have achieved unprecedented, sometimes super-human accuracy on many modern applications in artificial intelligence, such as image classification and speech recognition. Power dissipation is however a major concern in these energy-hungry machine-learning architectures, and decreasing it requires designs that provide a more energy-efficient combination of hardware and machine-learning algorithms. There is an increased emphasis to leverage parallelism and specialization to improve performance and energy efficiency. To dramatically reduce power consumption, silicon photonics has been proposed to improve performance-per-Watt compared to electrical implementation.This project leverages photonic technology and heterogeneous multicores for the design of deep-neural network accelerators that improve parallelism, concurrency, energy efficiency and scalability in various machine-learning applications. The first task of the project is concerned with the characterization and identification of photonic devices that can implement accelerator functionalities such as multiply-and-accumulate, summation, and other arithmetic operations. The characterized devices are then inserted into single-layer and multi-layer photonic topologies for implementing accelerator functionality. The second task of the project implements various types of deep-learning architectures on the proposed photonic neural network accelerator to maximize the gains offered by the photonic technology. The third task of the project builds an extensive simulation and modeling infrastructure that combines the photonic technology, network architectures, accelerator functionality, and machine-learning algorithms developed in the previous two steps, in order to validate the significant reduction in energy consumption enabled by the photonic neural-network accelerator.The proposed research bridges a very important gap between photonic technology, hardware architecture, and machine learning. As such, and due to its cross-cutting nature, it is expected to have far-reaching impacts on the design of next-generation multicore architectures. It will foster new research directions in several areas, spanning computer architecture, optical technology, machine learning algorithms and applications. The research will also play a major role in education by integrating discovery with teaching and training. All the research findings and simulation toolkits will be disseminated to the community via conference and journal publications, and a dedicated website.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.