Architecture and Circuit Design Optimization for Compute-In-Memory

The objective of the proposed research is to optimize computing-in-memory (CIM) design for accelerating Deep Neural Network (DNN) algorithms. As compute peripheries such as analog-to-digital converter (ADC) introduce significant overhead in CIM inference design, the research first focuses on the circuit optimization for inference acceleration and proposes a resistive random access memory (RRAM) based ADC-free in-memory compute scheme. We comprehensively explore the trade-offs involving different types of ADCs and investigate a new ADC design especially suited for the CIM, which performs the analog shift-add for multiple weight significance bits, improving the throughput and energy efficiency under similar area constraints. Furthermore, we prototype an ADC-free CIM inference chip design with a fully-analog data processing manner between sub-arrays, which can significantly improve the hardware performance over the conventional CIM designs and achieve near-software classification accuracy on ImageNet and CIFAR-10/-100 dataset. Secondly, the research focuses on hardware support for CIM on-chip training. To maximize hardware reuse of CIM weight stationary dataflow, we propose the CIM training architectures with the transpose weight mapping strategy. The cell design and periphery circuitry are modified to efficiently support bi-directional compute. A novel solution of signed number multiplication is also proposed to handle the negative input in backpropagation. Finally, we propose an SRAM-based CIM training architecture and comprehensively explore the system-level hardware performance for DNN on-chip training based on silicon measurement results.Ph.D