SRAM-Based Analog Compute-In-Memory Architecture Using C-2C Ladder And Signal Margin Assisted Design Methodology

Dinesh Kushwaha1, Ashish Joshi2, Abhishek Goel3, Rajiv Joshi4, Sudeb Dasgupta3, Anand Bulusu3
1Student, 2Intel India, 3IIT Roorkee, 4IBM TJ Watson


Abstract

In this manuscript, a signal margin-assisted design methodology for a current-based analog compute-in-memory (CIM) architecture is presented. Validation of the design methodology is done on the proposed analog compute-in-memory (CIM) architecture using a C-2C ladder. The effectiveness of an analog CIM architecture depends on the signal margin, scalability, throughput, energy efficiency, computation density, accuracy, and robustness. Hence, we proposed a design methodology to evaluate the effectiveness of an analog CIM architecture at the early design stage. The proposed design methodology can consider the device-level variations to estimate system-level performance and computation accuracy. An evaluation model is also developed to explore the proposed design methodology. LeNet-5 CNN is used to validate the proposed evaluation model. The signal margin of the proposed architecture is 35 mV, which is 2.5× higher than the state-of-the-art. The energy efficiency of the proposed architecture is 1666 TOPS/W at 0.9 V supply voltage in 28 nm CMOS technology. The inference accuracy for the MNIST data set is 98.6 %.