In the era of big data, conventional computers are increasinglychallenged by theapproaching end of Moore’s law, and by the latency/energy burdens of moving data between memory andprocessor in the von Neumann architecture.To get around these fundamental limits, in-memory computing (IMC) has recently emerged as a promising technique to conduct computing in situ, i.e., within the memory unit. One typical IMC methodology is based on resistive memories, such as resistive switching memory, phase change memory and magnetic memory.By exploiting the crosspoint memory architecture or the resistive switching dynamics, various IMC schemes can be realized.
Here we first showthat a crosspointarray of analog resistive memoriescan directly solve a system of linear equations, or find the matrix eigenvectors. Thesecomputations are completed in just one single step, thanks to the physical computing with Ohm’s and Kirchhoff’s laws, and thanks to the feedback connections in thecrosspointcircuit. Representativealgebraic problems are demonstrated in hardware and applied to classical computing tasks, such asimplementing Google’s PageRank and solving the Schr?dinger equation in one step. Then, we introducethe concept of stateful neural network, which performs allBooleanlogicoperations with the samecircuit topology of resistiveswitches. Stateful neural network can solve all2-input logic operations with just one step, except for the XOR needing two sequential steps. 1-bit full adder is realized with just two steps and five resistive switches, thus highlighting the high efficiencies of space, time, and energy of the stateful neural network for logic computing.
孫仲，2011畢業于南開大學材料物理系，獲理學學士學位；隨後進入清華大學物理系學習，于2016年獲理學博士學位。目前在意大利米蘭理工大學從事博士後研究工作。主要从事新型存储器及存内计算、机器學習芯片等方面的研究，在新型存储器物理机制，存内逻辑计算、模拟计算和机器學習等领域开展了多项开拓性研究，取得了多项创新成果。近三年的研究成果发表在PNAS、Nat. Commun.和Adv. Mater.等國際頂級期刊上，並得到國際媒體的廣泛關注和報道。