Abstract:
This investigation proposes a computing-in-memory (CIM) design to circumvent the von Neumann bottleneck which causes limited computation throughput for effective artifici...Show MoreMetadata
Abstract:
This investigation proposes a computing-in-memory (CIM) design to circumvent the von Neumann bottleneck which causes limited computation throughput for effective artificial intelligence (AI) applications. The proposed CIM performs multiple operations such as single-instruction basic Boolean operations, addition, and signed number multiplication, and multiple functions such as normal mode and retention mode for the built-in self-test (BIST). Its 2T-Switch requires only two transistors to be utilized for static random-access memory (SRAM) array; thus, the arithmetic unit can be chosen easily and the area overhead is minimized. Its ripple carry adder and multiplier (RCAM) unit based on single-ended disturb-free 7T 1-Kb SRAM was developed using the full swing-gate diffusion input (FS-GDI) technology that has full voltage swing resolution, low power consumption, and less chip area cost. Its Auto-Switching Write Back Circuit restores addition and multiplication operations automatically to assigned memory address. The CIM is implemented using the TSMC 40-nm CMOS process, where the core area is 432.81 \times 510.265\,\,\mu \text{m}^{2} . Among the related works, the proposed CIM performs the most number of operations and functions.
Published in: IEEE Transactions on Very Large Scale Integration (VLSI) Systems ( Volume: 29, Issue: 12, December 2021)