ANALYSIS OF THE ENERGY VS. ACCURACY TRADEOFF IN THE 10T1C BITCELL-BASED IN-MEMORY ARCHITECTURE

Ji, Shi

Permalink

https://hdl.handle.net/2142/124806 Copy

Description

Title

ANALYSIS OF THE ENERGY VS. ACCURACY TRADEOFF IN THE 10T1C BITCELL-BASED IN-MEMORY ARCHITECTURE

Author(s)

Ji, Shi

Issue Date

2023-05-01

Keyword(s)

In-Memory Computing; Deep In-Memory Architecture; circuit modeling;

Abstract

The enormous data movement between memory and processor has been an obstacle to the deployment of deep neural networks (DNNs) in edge applications, where stringent constraints on energy and latency are required. In-memory computing (IMC) has emerged as a promising candidate to solve this issue by embedding computation in memory. Many works have studied IMCs with different design choices; however, a systematic approach to optimizing the design of IMCs has not been thoroughly studied. This is due to the lack of understanding of the impact of architectural choices and circuit design parameters on system-level metrics, such as compute signal-to-noise ratio (SNR), energy efficiency, and compute density. In this work, we propose a circuit-aware Python model of accuracy and energy for a charge-redistribution (QR) based IMC architecture with a 10-transistor-1-capacitor (10T1C) cell structure. The model incorporates Cadence circuit simulation results and circuit behavior and is validated via Monte Carlo simulations in a 28nm CMOS technology. Through the model, we demonstrate that the compute SNR of 10T1C QR IMCs is limited by the interplay of critical design parameters, including the capacitor size of the 10T1C bit cell, analog-to-digital converter (ADC) precision, and ADC input noise. Our systematic modeling approach enables the determination of the effect of critical design parameters on system performance, encompassing both energy efficiency and accuracy. The behavioral model attains a normalized mean square error less than −23.61 𝑑𝐵, demonstrating its accuracy. The proposed model can be used to design QR IMC architectures for target applications with optimal performance or to produce accurate QR IMC behaviors with specific circuit parameters.

Type of Resource

text

Language

eng

Owning Collections