SamurAI: a 1.7MOPS-36GOPS Adaptive Versatile IoT Node with 15,000x Peak-to- Idle Power Reduction, 207ns Wake-up Time and 1.3TOPS/W ML Efficiency

Ivan Miro-Panades¹, Benoit Tain², Jean-Frédéric Christmann¹, David Coriat¹, Romain Lemaire¹, Clement Jany³, Baudouin Martineau³, Fabrice Chaix³, Anthony Quelen³, Emmanuel Pluchart¹, Jean-Philippe Noel¹, Reda Boumchedda^3,4, Adam Makosiej³, Maxime Montoya³, Simone Bacles-Min¹, David Briand², Jean-Marc Philippe², Alexandre Valentian¹, Frédéric Heitzmann³, Edith Beigne³, Fabien Clermidy^1
1 Univ. Grenoble Alpes, CEA, LIST, Grenoble, France;
² Univ. Paris-Saclay, CEA, LIST, Gif sur Yvette, France;
³ Univ. Grenoble Alpes, CEA, LETI, Grenoble, France;
⁴ STMicroelectronics, Crolles, France

Abstract

IoT node application requirements are torn between sporadic data-logging and energy-hungry data processing (e.g. image classification). This paper presents a versatile IoT node covering this gap in processing and energy by leveraging two on-chip sub-systems: a low power, clock-less, event-driven Always-Responsive (AR) part and an energy-efficient On- Demand (OD) part. The AR contains a 1.7MOPS event-driven, asynchronous Wake-up Controller (WuC) with 207ns wake-up time optimized for short sporadic computing. OD combines a deep-sleep RISC-V CPU and 1.3TOPS/W Machine Learning (ML) and crypto accelerators for more complex tasks. The node can perform up to 36GOPS while achieving 15,000x reduction from peak-to-idle power consumption. The interest of this versatile architecture is demonstrated with 105μW daily average power on an applicative classification scenario.

Introduction

An event-driven IoT node is a way to reduce the power consumption of sporadic computing. SamurAI (Fig. 1) combines an event-driven WuC using asynchronous logic (low-energy, clock-less, and fast wake-up time) in the AR subsystem with an energy efficient synchronous RISC-V CPU including specialized accelerators in the OD sub-system to make a versatile IoT node. Depending on the application needs, one or both cores can be used as shown in Fig. 2.

Fig. 1: SamurAI system architecture, with Always-Responsive and On-Demand sub-systems and associated power domains.

Fig. 2: SamurAI power modes.