eDRAM-CIM: Reconfigurable Charge Domain Compute-In-Memory Design With Embedded Dynamic Random Access Memory Array Realizing Adaptive Data Converters

Abstract

This article presents a compute-in-memory (CIM) architecture for a large-scale machine learning (ML) accelerator, which employs <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1T1C$</tex-math> </inline-formula> embedded dynamic random access memory (eDRAM) bitcells as charge domain circuits for convolution neural network (CNN) multiply-accumulation-averaging (MAV) computations. By repurposing existing <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1T1C$</tex-math> </inline-formula> eDRAM columns to construct an adaptive data converter, dot-product, averaging, pooling, and rectified linear unit (ReLU) activation on the memory array, the eDRAM-CIM design eliminates the need for an extra dedicated hardware accelerator, which significantly reduces the hardware implementation cost and increases the reconfigurability of the CIM computation circuit. In the eDRAM-CIM design, 8 b digital inputs from image pixel values are converted to analog domain directly on eDRAM columns. The dot-product is computed without disturbing the kernel weights, which are stored in the eDRAM array, preventing data duplication and additional intra-memory data movement. In addition, the convolution results are transferred back to the digital domain using an in-eDRAM adaptive dynamic range successive-approximation (SAR) analog-to-digital converter (ADC) utilizing a narrow range of dot-product distribution to reduce ADC latency and energy. A 16 Kb eDRAM-CIM prototype, implemented in the 65 nm CMOS process, demonstrates the concept and the functionality of the test-chip using the CIFAR-10 dataset with 8 b input and 8 b signed weight, achieving 90.77% accuracy, 4.71 GOPS throughput, 4.76 TOPS/W energy efficiency, and 8.26 GOPS/mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> . A scalability analysis is developed to show that the presented eDRAM-CIM approach, which was adopted in an advanced eDRAM technology node, shows promising energy efficiency (11.21 TOPS/W) and throughput (22.2 TOPS), suggesting its potential for application in large-scale energy-efficient CIM designs.

Topics & Concepts