ABSTRACT
Artificial intelligence (AI) edge devices prefer employing high-capacity nonvolatile compute-in-memory (CIM) to achieve high energy efficiency and rapid wakeup-to-response with sufficient accuracy. Most previous works are based on either memristor-based CIMs, which suffer from accuracy loss and do not support training as a result of limited endurance, or digital static random-access memory (SRAM)-based CIMs, which suffer from large area requirements and volatile storage. We report an AI edge processor that uses a memristor-SRAM CIM-fusion scheme to simultaneously exploit the high accuracy of the digital SRAM CIM and the high energy-efficiency and storage density of the resistive random-access memory memristor CIM. This also enables adaptive local training to accommodate personalized characterization and user environment. The fusion processor achieved high CIM capacity, short wakeup-to-response latency (392 microseconds), high peak energy efficiency (77.64 teraoperations per second per watt), and robust accuracy (<0.5% accuracy loss). This work demonstrates that memristor technology has moved beyond in-lab development stages and now has manufacturability for AI edge processors.
ABSTRACT
Artificial Intelligence (AI) is currently experiencing a bloom driven by deep learning (DL) techniques, which rely on networks of connected simple computing units operating in parallel. The low communication bandwidth between memory and processing units in conventional von Neumann machines does not support the requirements of emerging applications that rely extensively on large sets of data. More recent computing paradigms, such as high parallelization and near-memory computing, help alleviate the data communication bottleneck to some extent, but paradigm- shifting concepts are required. Memristors, a novel beyond-complementary metal-oxide-semiconductor (CMOS) technology, are a promising choice for memory devices due to their unique intrinsic device-level properties, enabling both storing and computing with a small, massively-parallel footprint at low power. Theoretically, this directly translates to a major boost in energy efficiency and computational throughput, but various practical challenges remain. In this work we review the latest efforts for achieving hardware-based memristive artificial neural networks (ANNs), describing with detail the working principia of each block and the different design alternatives with their own advantages and disadvantages, as well as the tools required for accurate estimation of performance metrics. Ultimately, we aim to provide a comprehensive protocol of the materials and methods involved in memristive neural networks to those aiming to start working in this field and the experts looking for a holistic approach.