A Photonic-Electronic non-von Neumann Processor Core for Highly Efficient Computing (APT-NuCOM)

现代社会在很大程度上依赖于大量数据的生成、处理和传输。据预测，到2025年，地球仪将产生175 zettabytes（175万亿千兆字节）的数据，而所谓的“边缘计算”设备将产生超过90 zettabytes。处理如此大量的数据需要不断增加的计算能力、存储器和通信带宽-这些需求无法通过传统的数字电子技术来持续满足。如果我们考虑训练各种人工智能（AI）系统所需的计算总数（以1 PetaFLOP/s的速率计算的#天为单位）的历史趋势，那么当今信息技术的需求和能力之间日益扩大的差距就是一个例证。这一趋势遵循摩尔定律（大约每两年翻一番），直到2012年，之后翻一番的时间减少到仅仅3.4个月！这一趋势因库米定律的崩溃而加剧，该定律指出，每焦耳能量的计算次数大约每1.5年翻一番。直到最近，这个定律也一直被遵循，但我们现在正在接近CMOS电子器件和冯-诺依曼架构的广泛接受的计算效率墙，大约为10 GMAC/焦耳（MAC是一种乘法累加运算）。因此，用于训练现代人工智能系统的能源消耗确实令人震惊，从而对可持续性产生不利影响。这导致AI从标准的CPU设计转向使用具有上级并行性的协处理器-GPU、ASIC、FPGA。然而，即使在这里，电信号的限制也会导致大量的能源消耗。例如，最近估计，用于精确机器翻译的大型基于GPU的自然语言处理系统的训练导致的二氧化碳排放量相当于5辆汽车的终身使用！显然，需要一种新的方法。因此，在APT-NuCOM项目中，我们将开发一种高效的新型非冯·诺依曼协处理器，该协处理器利用光子计算提供的明显优势，但同时与电子领域无缝链接，以实现与现有电子计算基础设施的集成。APT-NuCOM协处理器将利用新的相变光子内存计算概念，以PetaMAC/s的速度提供大规模并行计算，最终达到接近人脑的能量预算。

Modern society depends massively on the generation, processing and transmission of vast amounts of data. It is predicted that by 2025, 175 zettabytes (175 trillion gigabytes) of data will be generated around the globe, with so-called 'edge computing' devices creating more than 90 zettabytes alone. Processing such huge amounts of data demands ever increasing computational power, memory and communication bandwidth - demands that cannot be sustainably met by conventional digital electronic technologies. The growing gap between the needs and the capabilities of today's information technology is exemplified if we consider the historical trend in total number of computations (in units of #days of calculating at a rate of 1 PetaFLOP/s) needed to train various artificial intelligence (AI) systems. The trend followed Moore's Law (doubling approximately every two years) until 2012, after which the doubling time reduced to a mere 3.4 months! This trend is compounded by the breakdown in Koomey's Law, which states that the number of computations per Joule of energy doubles around every 1.5 years. This law was also followed until quite recently, but we are now approaching a widely accepted computing efficiency-wall at around 10 GMAC/Joule (a MAC is a multiply-accumulate operation) for CMOS electronics and the von-Neumann architecture. As a result, the energy consumption used in training modern AI systems is truly staggering, with consequent adverse effects for sustainability. This has led to a move away from standard CPU designs in AI towards the use of co-processors - GPUs, ASICs, FPGAs - with superior parallelism.However, even here the limitations of electrical signalling lead to massive levels of energy consumption. It was recently estimated, for example, that the training of a large GPU-based natural language processing system used for accurate machine translation resulted in carbon dioxide emissions equivalent to lifetime use of 5 cars! Clearly, a new approach is needed. Thus, in the APT-NuCOM project we will develop a highly efficient novel non-von Neumann co-processor that exploits clear advantages offered by photonic computation, but at the same time links seamlessly with the electronic domain to enable integration with existing electronic computing infrastructure. The APT-NuCOM co-processor will exploit novel phase-change photonic in-memory computing concepts to deliver massively parallel computation at PetaMAC/s speeds and, ultimately, an energy budget approaching that of the human brain.