Analysis of vertical AND flash memory for energy-efficient, scalable, fast CIM beyond vertical NAND flash memory

Area of Science:

• Semiconductor Device Physics
• Computer Architecture
• Memory Technologies

Background:

• The von Neumann architecture faces a data movement bottleneck, hindering performance and energy efficiency.
• Compute-in-memory (CIM) architectures are proposed to address this bottleneck by performing computation within memory.
• Vertical NAND (V-NAND) flash memory is a candidate for CIM, but faces limitations in energy efficiency due to structural constraints.

Purpose of the Study:

• To conduct a comprehensive comparison between V-NAND and vertical AND (V-AND) flash memory for CIM applications.
• To evaluate the energy efficiency of V-AND and V-NAND in different CIM operational regimes.
• To identify the key factors contributing to energy efficiency differences between V-AND and V-NAND.

Main Methods:

• Development of analytical models to characterize the performance and energy consumption of V-NAND and V-AND for CIM.
• Experimental validation of the analytical models using fabricated memory devices.
• Comparative analysis of energy efficiency under varying inference counts and stack heights.

Main Results:

• V-AND demonstrates superior energy efficiency compared to V-NAND in CIM applications.
• The energy efficiency advantage of V-AND is more pronounced in low-inference-count scenarios and with increased memory stack height.
• Elimination of the bias pass requirement in V-AND significantly contributes to its improved energy efficiency.

Conclusions:

• V-AND flash memory presents a more energy-efficient solution for CIM accelerators than V-NAND.
• V-AND's structural advantages make it a promising technology for developing scalable and fast CIM systems.
• Further research into V-AND based CIM is warranted to fully exploit its potential for next-generation computing.