A 65-nm Digital Stochastic Compute-in-Memory CNN Processor With 8-bit Precision

Compute-in-memory (CIM) is an emerging solution that embeds compute logic inside the memory array to achieve high energy efficiency and reduce memory access cost. However, current CIM solutions require bulky analog-to-digital converter (ADC)/digital-to-analog converters (DACs) that make the accelerator power hungry and unable to scale in the more advanced process with a low power supply. Stochastic computing (SC) is an attractive alternative digital compute scheme that uses tiny bit-serial logic, e.g., and and or, to perform the basic MAC operations on stochastic sequences. In this work, we propose a stochastic CIM (SCIM) approach that combines the benefits of SC and CIM accelerators. It eliminates the DACs/ADCs in CIM and adds CIM’s benefits to SC. We demonstrate a highly programmable convolutional neural network (CNN) accelerator that achieves up to 7.96-TOPS/W peak energy efficiency in 65 nm and inference accuracy comparable to 8-b fixed point.