H2020Индивидуална стипендия2016–2018

SelectX · Integrated Crossbar of Microelectromechanical Selectors and Non-Volatile Memory Devices for Neuromorphic Computing

„Хоризонт 2020“ — Действия „Мария Склодовска-Кюри“

Период: 2016-04-01 → 2018-03-31
Финансиране от ЕС: 125 423 €
Участници: 1
Схема: MSCA-IF-EF-RI

Линиите свързват координатора с партньорите.

Накратко на български

Мемристорите и микромеханичните превключватели се комбинират, за да имитират работата на синапсите в мозъка. Това помага за създаването на по-енергоефективен хардуер за изкуствени невронни мрежи с висока плътност на данните.

Този кратък обзор е генериран от изкуствен интелект

Кратко обяснение, генерирано от езиков модел по текста на CORDIS. Оригиналът е по-долу.

Резултати накратко

Integrated Crossbar of Microelectromechanical Selectors and Non-Volatile Memory Devices for Neuromorphic Computing

Manufacturing of ANNs is currently limited, since traditional CMOS hardware cannot provide the high synaptic density of biological neural circuits with reasonable energy consumption. A promising solution is to combine transistor hardware with an alternative technology, like the emerging memristors (or ReRAM), that can deliver the artificial synaptic functionality in an energy efficient way. A memristor has an analog tunable resistance that mimics the basic synaptic functionality, which would otherwise take dozen transistors to reproduce. A memristor stores data in a non-volatile fashion without drawing power and has only two terminals, making it energy efficient, scalable and stackable. Crossbars of memristors stacked on transistor chips (Fig 1a) could deliver the high density and connectivity required for hardware ANNs. However, the crossbar architecture has the sneak path problem: neighbouring devices create an electrical short around the selected device (Fig 1b). A solution is to use a non-linear selector for each of the memory devices, the most common selector being a transistor. The transistor is a three-terminal device fabricated in the Silicon bulk so it has limited scalability and stackability. Therefore its use as a selector negates the advantages offered by the two-terminal memristor. This project proposed an alternative selector based on a two-terminal microelectromechanical switch (MEMS) (Fig. 2), thanks to its high nonlinearity.The goal was to develop a crossbar of memristors integrated with MEMS selectors. The success of this project is immediately relevant to hardware ANNs and non-volatile memories and to the industry players in the field (HP, IBM, etc.).

Текст от CORDIS, на английски · Данни: CORDIS, © Европейски съюз

Цел на проекта

Building artificial neuronal networks (ANN) that mimic their biological prototypes is one of the remaining grand challenges in computing. Despite transistor scaling and improved architectures, modern supercomputers require kWs of power to replicate functionality for which living neural networks take only a few Ws. Currently there is no hardware that can provide a high synaptic density with reasonable energy consumption. Hybrid CMOS chips with stacked crossbars of analog non-volatile memory devices (NVM), like memristors, promise to deliver the required high density and connectivity. However, the crossbar architecture suffers from the sneak path problem, neighbouring devices creating electrical shorts around the selected device. Biological systems do not have this problem due to the inherent nonlinearity of their potentiation and spiking. A high nonlinearity can be reproduced in a crossbar by using a selector for each memory device. The selector commonly used is a transistor which limits the scalability and stackability. This project proposes an alternative selector based on a two-terminal MEMS switch. MEMS switches are heavily researched for RF applications, but their high nonlinearity makes them attractive as selectors for NVM. This project will design, simulate and fabricate a crossbar of integrated MEMS selectors and TiO2 memristor devices. Initially, the selectors and memristors will be optimized separately, then monolithically integrated. Their scalability and stackability will be investigated. Finally, a prototype 3x3 crossbar of integrated memristor/MEMS selectors will be fabricated and used to demonstrate vector matrix multiplication - a foundational element of many complex ANNs like a perceptron. The proposed work is not linked to a particular NVM (ReRAM is an example), being suitable for any dense crossbar system that requires selectors. The findings are relevant to the fields of hardware ANNs and non-volatile memories and to major industry players.

Оригинален текст от CORDIS (на английски).

Участници

INSTITUTUL NATIONAL DE CERCETAREDEZVOLTARE PENTRU MICROTEHNOLOGIE · VOLUNTARIКоординаторРумъния