基于自旋磁存储器的存内处理技术研究

The challenges of big-data processing and the bottlenecks of transitional von-Neumann computing architecture have become one key contradiction in the current information and computer research areas. The data-centric new computing architectures and the corresponding processing-in-memory technologies have been widely recognized as potential solutions. The objective of the processing-in-memory technology is to build an integrated computing platform that can store and process data in a single memory chip. In the current stage, however, designing an intelligent memory chip is the key to effectively implement the processing-in-memory technology. Recently, emerging nonvolatile memories have brought a new hope. In this project, we aim to develop an efficient processing-in-memory solution based on spintronic memories, from a cross-layer perspective in terms of computing paradigm, circuit design, and computing architecture. First of all, we will design a novel computing paradigm suitable for spintronic memories, establish a complete set of logic functions, and implement the corresponding control scheme. Then, we will design and optimize the spintronic memory circuits so as to support both data storage and data processing functions within the same chip, which is also expected to support dynamic re-configurability between the storage capacity and processing capacity. Finally, we will develop a master-slave heterogeneous computing architecture that can support efficient processing-in-memory capability. We believe that this project has important strategic significance and economic value to promote the development of the information industry in our country.

大数据处理的挑战与冯诺依曼计算架构的瓶颈成为当前信息领域的关键矛盾之一。以数据为中心的新型计算架构以及相应的存内处理技术受到人们的广泛认可。存内处理技术的目标是构建一个存算一体化软硬件平台，内存本身既能存储数据，也能处理数据。目前，智能内存芯片成为制约存内处理技术发展的关键。近年来，新型非易失性存储器为存内处理技术的高效实施带来了曙光。本项目研究基于自旋磁存储器的存内处理技术，从计算范式、芯片电路、计算架构三个层面开展跨层联合研究，旨在实现一个高效的存内处理技术方案。首先，构建适用于自旋磁存储器的存内处理计算范式、实现完备的逻辑计算功能集、设计相应的存内处理核与控制方案；然后，设计同时支持数据存储与数据处理功能的自旋磁存储器芯片电路原型，且支持存储能力与处理能力的动态可重构；最后，设计能够支撑高效存内处理技术的主从异构计算架构。本项目对促进我国信息产业的发展具有重要的战略意义与经济价值。

随着大数据、物联网、人工智能等应用的快速兴起，数据以爆发式的速度进行增长，智能大数据处理挑战与冯诺依曼计算架构瓶颈成为当前信息领域的关键矛盾之一。此外，随着器件尺寸的进一步微缩，制造成本急剧增加，单纯依靠先进工艺进一步提升芯片性能的技术路径难以为继。在此背景下，以数据为中心的新型计算架构，如存内处理技术，受到人们的广泛认可与重点研究。存内处理技术的目标是构建一个存算一体化软硬件平台，内存本身既能存储数据，也能处理数据，并可支持动态重构，从而大幅减小数据搬运，降低能耗。目前，智能内存芯片成为制约存内处理技术发展的关键。近年来，新型非易失性存储器介质为存内处理技术的高效实施带来了曙光，例如自旋磁存储器、忆阻器、相变存储器以及铁电存储器等。本项目研究基于自旋磁存储器介质的存内处理技术，重点从计算范式、芯片电路、计算架构三个层面开展跨层联合研究，旨在实现一个高效的自旋存内处理技术方案。首先，构建了一个适用于自旋磁存储器介质的存内处理计算范式，通过读写方式与外围电路的改造，可通过存储单元实现完备的逻辑计算功能集，并设计了相应的存内处理核与控制方案；然后，设计了同时支持数据存储与数据处理功能的自旋磁存储器芯片电路原型，且支持存储能力与处理能力的动态可重构，从而实现计算能力与存储能力的动态均衡；最后，基于RISC-V指令集，设计了能够支撑高效存内处理技术的主从异构计算架构，能够支持任务调度与流水线处理，从而实现高效的计算与控制。本项目对我国存内计算芯片的发展具有重要的理论意义与实际价值。