A full spectrum of computing-in-memory technologies
A full spectrum of computing-in-memory technologies
复制标题
DOI：
10.1038/s41928-023-01053-4
复制
发表时间：
2023-11
影响因子：
34.3
通讯作者：
Zhong Sun;Shahar Kvatinsky;Xin Si;Adnan Mehonic;Yimao Cai;Ru Huang
中科院分区：
文献类型：
--
作者：
Zhong Sun;Shahar Kvatinsky;Xin Si;Adnan Mehonic;Yimao Cai;Ru Huang
内存计算（CIM）可以用来克服冯诺依曼瓶颈，并提供可持续的提高计算吞吐量和能源效率。不同CIM方案的基础是两种计算原语的实现：逻辑门和乘-累加运算。考虑到任一操作中的输入和输出，CIM技术在存储单元如何参与计算过程方面有所不同。这种复杂性使得建立对CIM技术的全面理解变得困难。在这里，我们通过识别参与计算的存储单元作为输入和/或输出的程度，提供了所有CIM技术的全谱分类。我们阐明了标准CIM技术在这个频谱的详细原则，并提供了一个平台，比较每个不同的技术的优点和缺点。我们的分类法也可以潜在地用于开发其他CIM方案，通过将频谱应用于不同的存储设备和计算原语。
Computing in memory (CIM) could be used to overcome the von Neumann bottleneck and to provide sustainable improvements in computing throughput and energy efficiency. Underlying the different CIM schemes is the implementation of two kinds of computing primitive: logic gates and multiply–accumulate operations. Considering the input and output in either operation, CIM technologies differ in regard to how memory cells participate in the computation process. This complexity makes it difficult to build a comprehensive understanding of CIM technologies. Here, we provide a full-spectrum classification of all CIM technologies by identifying the degree of memory cells participating in the computation as inputs and/or output. We elucidate detailed principles for standard CIM technologies across this spectrum, and provide a platform for comparing the advantages and disadvantages of each of the different technologies. Our taxonomy could also potentially be used to develop other CIM schemes by applying the spectrum to different memory devices and computing primitives.