RESUMEN
The existing von Neumann architecture for artificial intelligence (AI) computations suffers from excessive power consumption and memory bottlenecks. As an alternative, compute-in-memory (CIM) technology has been emerging. Among various CIM device candidates, split-gate NOR flash offers advantages such as a high density and low on-state current, enabling low-power operation, and benefiting from a high level of technological maturity. To achieve high energy efficiency and high accuracy in CIM inference chips, it is necessary to optimize device design by targeting low power consumption at the device level and surpassing baseline accuracy at the system level. In split-gate NOR flash, significant factors that can cause CIM inference accuracy drop are the device conductance variation, caused by floating gate charge variation, and a low on-off current ratio. Conductance variation generally has a trade-off relationship with the on-current, which greatly affects CIM dynamic power consumption. In this paper, we propose strategies for designing optimal devices by adjusting oxide thickness and other structural parameters. As a result of setting Tox,FG to 13.4 nm, TIPO to 4.6 nm and setting other parameters to optimal points, the design achieves erase on-current below 2 µA, program on-current below 10 pA, and off-current below 1 pA, while maintaining an inference accuracy of over 92%.
RESUMEN
In recent years, compute-in-memory (CIM) has been extensively studied to improve the energy efficiency of computing by reducing data movement. At present, CIM is frequently used in data-intensive computing. Data-intensive computing applications, such as all kinds of neural networks (NNs) in machine learning (ML), are regarded as 'soft' computing tasks. The 'soft' computing tasks are computations that can tolerate low computing precision with little accuracy degradation. However, 'hard' tasks aimed at numerical computations require high-precision computing and are also accompanied by energy efficiency problems. Numerical computations exist in lots of applications, including partial differential equations (PDEs) and large-scale matrix multiplication. Therefore, it is necessary to study CIM for numerical computations. This article reviews the recent developments of CIM for numerical computations. The different kinds of numerical methods solving partial differential equations and the transformation of matrixes are deduced in detail. This paper also discusses the iterative computation of a large-scale matrix, which tremendously affects the efficiency of numerical computations. The working procedure of the ReRAM-based partial differential equation solver is emphatically introduced. Moreover, other PDEs solvers, and other research about CIM for numerical computations, are also summarized. Finally, prospects and the future of CIM for numerical computations with high accuracy are discussed.