Functionality Enhanced Memories for Edge-AI Embedded Systems

Abstract

With the surge in complexity of edge workloads, it appeared in the scientific community that such workloads cannot be anymore overflown to the cloud due to the huge edge device to server communication energy cost and the high energy consumption induced in high end server infrastructure. In this context, edge devices must be able to efficiently process complex data-intensive workloads bringing in the concept of Edge AI. However, current architectures show poor energy efficiency while running data intensive workloads. While the community looks toward the integration of new memory architectures using emerging resistive memories and new specific accelerators, we propose a new concept to boost the energy efficiency of Edge systems running data intensive workloads : Functionality Enhanced Memories (FEM). FEM consist on a memory architecture with new functionalities at a decent area overhead cost. In this work, we demonstrate the feasibility of native transpose access for 1Transistor-1RRAM bitcells leveraging three independent gates transistors. Based on that, we thereby propose a concept of FEM-enabled Edge system embedding the proposed native transpose access RRAM-based memory architecture and an in-SRAM computing architecture (the BLADE).

Artificial Intelligence in the IoT Era: A Review of Edge AI Hardware and Software

“…Baller et al measured five edge devices and give their recommendations for best performance in continuous and sporadic scenarios [34]. Operating AI inference in industrial conditions could be made more robust by using magnetoresistive random access memory (MRAM) [35], [36] Energy efficiency is a constant concern with Edge AI, and there are developments in this area, such as Levisse et al with their functionality enhanced memories [37]. Liu et al propose hybrid parallelism, which makes hierarchical training of AI models for Edge AI situations efficient [38].…”

Write Termination Circuits for RRAM: A Holistic Approach From Technology to Application Considerations

“…Now that it is clear that the current cloud-based infrastructure is not scalable and cannot sustain the rise of deported dataintensive and machine learning workloads [1], new solutions have to be found to be able to run locally these applications. While new computing architectures and accelerators are being explored to cope with the new computing paradigm opened by machine learning workloads [2], [10], [11], [17]- [19], there is no clear solution on how to store the data needed by the application. As a reference, recent Convolutional Neural Networks (CNN) require from few MegaBytes (MB) to hundred of MB [20] of memory to run.…”

CNN Based Animal Repelling Device for Crop Protection

“…Animals and people will not be harmed by the embedded development system. The project's purpose is to develop a smart security system for agricultural safety [4][5][6][7]. When compared to other wildlife monitoring methods, camera traps are the most successful and cost-effective option for a variety of species [8].…”