Wireless Memory Approximation for Energy-efficient Task-specific IoT Data Retrieval

TL;DR

The paper addresses energy waste from DRAM standby refresh in resource-constrained IoT devices by introducing wireless memory activation and wireless memory approximation for task-specific data retrieval. It develops a pull-based framework that remotely activates memory and configures refresh periods based on model importance, supported by analytical models of retention and retrieval accuracy and a decomposed energy model. Through grid-search optimization and Monte Carlo simulations, the authors demonstrate substantial energy savings under retrieval-accuracy constraints compared with an always-on baseline, validating the approach across diverse system settings. This work enables energy-efficient, latency-conscious memory management for edge AI in next-generation networks and points to future work on multi-task remote memory protocols.

Abstract

The use of Dynamic Random Access Memory (DRAM) for storing Machine Learning (ML) models plays a critical role in accelerating ML inference tasks in the next generation of communication systems. However, periodic refreshment of DRAM results in wasteful energy consumption during standby periods, which is significant for resource-constrained Internet of Things (IoT) devices. To solve this problem, this work advocates two novel approaches: 1) wireless memory activation and 2) wireless memory approximation. These enable the wireless devices to efficiently manage the available memory by considering the timing aspects and relevance of ML model usage; hence, reducing the overall energy consumption. Numerical results show that our proposed scheme can realize smaller energy consumption than the always-on approach while satisfying the retrieval accuracy constraint.

Figures (6)

• Figure 1: The overview of the system integrating remote memory activation/approximation.
• Figure 2: Analytical model.
• Figure 3: Retention accuracy defined in Eq. \ref{['eq:retention_accuracy']} as a function of memory activation period $Q$ (left figure) and $\Delta$ (right figure) for the results obtained by Eq. \ref{['eq:retention_accuracy']} (line) and by simulations (marker).
• Figure 4: Retrieval accuracy and normalized energy consumption obtained by analysis (line) and simulations (marker) against $\zeta$.
• Figure 5: The set of retrieval accuracy and normalized energy consumption for the different value of $\beta$ where we vary the value of $\tau$.