High performance and energy efficient deep learning processor design

随着传感器无处不在，物联网将引发大数据的大量流入。传感器数量的惊人增长将在未来几年产生更大量的数据。软件利用了一种被称为深度学习的人工智能技术，它使用模拟的神经元和突触来处理数据。近年来，深度学习在处理图像和音频等大数据方面取得了巨大进步。深度学习系统在各个领域都很有前途，包括生物医学/健康信息学、自动驾驶汽车、计算机视觉、自然语言处理和大数据分析应用。在许多这些应用中，深度学习系统可以达到接近甚至超过人类的性能。 虽然深度学习系统功能强大，但实现算法的成本非常昂贵。深度学习算法是计算和内存密集型的。为深度学习解决方案训练数据集需要大量数据。要执行任务以解决真实的世界问题，机器需要配备足够的处理能力。为了确保更高的效率和更少的时间消耗，需要多核高性能GPU和类似的处理单元。这些处理单元是昂贵的并且消耗大量功率。工业级深度学习系统需要高端数据中心，而无人机、机器人和其他移动的设备等智能设备需要小型但高效的处理单元。因此，在真实的世界中部署深度学习解决方案变得昂贵且耗电。 在拟议的研究计划中，将研究深度学习的三个方面，以应对上述挑战。首先，深度学习算法已被证明具有容错能力。近似计算是错误弹性深度学习算法的一个很好的候选者。其次，可以以低精度执行推理操作，并且可以以高精度处理训练过程。多精度实现有助于有效地支持具有多精度要求的操作。第三，高带宽存储器（HBM）结合存储器内处理（PIM）可以通过缓解存储器瓶颈和降低功耗来解决存储器带宽挑战。 在这项研究计划中，将通过提出近似多精度运算单元来构建基本处理单元，并采用HBM结合PIM，研究深度学习应用的新型硬件处理器架构的设计。此外，将研究基于近似计算的方法，以支持灵活性，稀疏性和训练。提出的深度学习处理器的主要目标是解决成本-能源-性能问题。该研究项目的成果将显著提高深度学习系统的性能和能效。

As sensors spread everywhere, the internet of things is going to trigger a massive inflow of big data. Incredible growth in the number of sensors will contribute to even larger amounts of data in the coming years. Software exploits an artificial intelligence technique known as deep learning, which uses simulated neurons and synapses to process data. Deep learning has produced dramatic advances in processing big data like images and audio in recent years. Deep learning systems are promising in various areas including biomedical/health informatics, self-driving cars, computer vision, natural language processing, and big data analysis applications. In many of these applications, deep learning systems can achieve near or even above human performance. Although deep learning systems are powerful, the cost to realize the algorithms is very expensive. Deep learning algorithms are computation and memory intensive. Training a data set for a deep learning solution requires massive data. To perform a task to solve real world problems, the machine needs to be equipped with adequate processing power. To ensure better efficiency and less time consumption, multi-core high performing GPUs and similar processing units are required. These processing units are costly and consume a lot of power. Industry level deep learning systems require high-end data centres and smart devices such as drones, robots and other mobile devices require small but efficient processing units. Thus, deploying deep learning solutions in the real world becomes a costly and power consuming affair. In the proposed research program, three aspects of deep learning will be investigated to address the aforementioned challenges. Firstly, deep learning algorithms have proven to be error-resilient. Approximate computing is a great candidate for error-resilient deep learning algorithms. Secondly, inference operations can be performed with low precision and training processes can be handled with high precision. Multiple-precision implementation is helpful to efficiently support operations with multiple-precision requirements. Thirdly, high bandwidth memory (HBM) combined with processing-in-memory (PIM) can address the memory bandwidth challenge by relieving the memory bottleneck and reducing power consumption. In this research program, the design of novel hardware processor architecture for deep learning applications will be investigated by proposing approximate multiple-precision arithmetic units to build basic processing elements and adopting HBM combined with PIM. In addition, approximate computing based methods will be investigated to support flexibility, sparsity, and training. The primary target of the proposed deep learning processor is to address cost-energy-performance issues. The outcome of this research program will significantly improve the performance and power efficiency of deep learning systems.