{"id":12934,"date":"2017-04-06T11:11:12","date_gmt":"2017-04-06T09:11:12","guid":{"rendered":"https:\/\/www.iemn.fr\/articles-temporaires-anglais\/frontiers-in-neuroscience-artificial-neuron-in-65-nm-cmos-technology-2.html"},"modified":"2019-10-10T14:57:56","modified_gmt":"2019-10-10T12:57:56","slug":"frontiers-in-neuroscience-artificial-neuron-in-65-nm-cmos-technology","status":"publish","type":"post","link":"https:\/\/www.iemn.fr\/en\/news\/frontiers-in-neuroscience-artificial-neuron-in-65-nm-cmos-technology.html","title":{"rendered":"Un neurone artificiel mille fois plus \u00e9conome en \u00e9nergie qu\u2019un neurone biologique"},"content":{"rendered":"<p><strong><a href=\"https:\/\/www.iemn.fr\/wp-content\/uploads\/2017\/04\/FIN.gif\"><img loading=\"lazy\" decoding=\"async\" class=\"alignright wp-image-12922\" src=\"https:\/\/www.iemn.fr\/wp-content\/uploads\/2017\/04\/FIN.gif\" alt=\"\" width=\"292\" height=\"98\" \/><\/a><\/strong><\/p>\n<blockquote><p><span style=\"color: #000000;\">Chef d\u2019\u0153uvre de l\u2019\u00e9volution, le cerveau humain est une source d\u2019inspiration pour les scientifiques. Des chercheurs de l\u2019IEMN et de l\u2019Ircica ont ainsi mis au point un neurone artificiel ultra-efficace en \u00e9nergie et qui reproduit tr\u00e8s pr\u00e9cis\u00e9ment les signaux \u00e9lectriques g\u00e9n\u00e9r\u00e9s dans le cerveau. Ces travaux sont publi\u00e9s dans la revue <em>Frontiers in Neuroscience<\/em>.<\/span><\/p><\/blockquote>\n<p>Dans notre cerveau, les neurones sont connect\u00e9s entre eux et g\u00e9n\u00e8rent une r\u00e9ponse binaire aux informations qu\u2019ils re\u00e7oivent des autres cellules nerveuses : soit ils \u00e9mettent un signal \u00e9lectrique, appel\u00e9 aussi potentiel d\u2019action, soit ils restent silencieux. Ce syst\u00e8me est \u00e0 la base de tous nos processus cognitifs et moteurs. Des chercheurs de l\u2019Institut d\u2019\u00e9lectronique, de micro\u00e9lectronique et de nanotechnologie (<a href=\"https:\/\/www.iemn.fr\/en\/\" target=\"_blank\" rel=\"noopener noreferrer\">IEMN<\/a>, CNRS\/Universit\u00e9 Lille I\/ISEN Lille\/Universit\u00e9 Valenciennes Hainaut-Cambresis\/\u00c9cole centrale de Lille) et de l\u2019Institut de recherche sur les composants logiciels et mat\u00e9riels pour l\u2019information et la communication avanc\u00e9e (<a href=\"http:\/\/www.ircica.univ-lille1.fr\/\" target=\"_blank\" rel=\"noopener noreferrer\">Ircica<\/a>, CNRS\/Universit\u00e9 Lille 1) ont reproduit ces propri\u00e9t\u00e9s \u00e0 l\u2019aide de dispositifs \u00e9lectroniques nanom\u00e9triques.<\/p>\n<p>Mesurant quelques microns carr\u00e9s, ces neurones artificiels sont dispos\u00e9s en grand nombre sur un circuit int\u00e9gr\u00e9 en silicium. Ils ne consomment que quelques dizaines de femtojoules (10-15 J) lors de la g\u00e9n\u00e9ration d\u2019un potentiel d\u2019action. Une performance \u00e9nerg\u00e9tique environ 1000 fois sup\u00e9rieure \u00e0 celle d\u2019un neurone biologique, et qui d\u00e9passe de plusieurs ordres de grandeur celle de tous les autres neurones artificiels existants. Ces travaux ouvrent de nombreuses perspectives, comme la cr\u00e9ation de r\u00e9seaux ultra-faible \u00e9nergie pour l\u2019intelligence artificielle. Ils pourraient \u00e9galement servir \u00e0 d\u00e9velopper les interactions entre neurones artificiels et neurones vivants, par exemple pour traiter des maladies comme celle de Parkinson ou r\u00e9parer des alt\u00e9rations de la moelle \u00e9pini\u00e8re. Cette \u00e9tude remet au passage en cause l\u2019id\u00e9e que les neurones naturels sont parfaitement optimis\u00e9s sur le plan \u00e9nerg\u00e9tique.<\/p>\n<p><a href=\"https:\/\/www.iemn.fr\/wp-content\/uploads\/2017\/04\/neurone_artificiel.gif\"><img loading=\"lazy\" decoding=\"async\" class=\"wp-image-13009 aligncenter\" src=\"https:\/\/www.iemn.fr\/wp-content\/uploads\/2017\/04\/neurone_artificiel.gif\" alt=\"\" width=\"698\" height=\"355\" \/><\/a><\/p>\n<p><strong>Contacts chercheurs :<\/strong><br \/>\n<a href=\"mailto:alain.cappy@ircica.univ-lille1.fr\">Alain Cappy<\/a> \u2013 IEMN et Ircica<br \/>\n<strong>Contact communication INSIS<\/strong> :<br \/>\n<a href=\"mailto:chloe.rimailho@cnrs-dir.fr\">insis.communication@cnrs.fr<\/a><\/p>\n<h6><span style=\"color: #000000;\">References :<\/span><\/h6>\n<p><span style=\"color: #333333;\"><strong>A 4-fJ\/Spike Artificial Neuron in 65 nm CMOS Technology<\/strong><\/span><br \/>\nIlias Sourikopoulos, Sara Hedayat, Christophe Loyez, Fran\u00e7ois Danneville, Virginie Hoel, Eric Mercier and Alain Cappy<br \/>\n<em>Front. Neurosci<\/em>., 15 March 2017<br \/>\n<a href=\"https:\/\/doi.org\/10.3389\/fnins.2017.00123\">https:\/\/doi.org\/10.3389\/fnins.2017.00123<\/a><\/p>\n<h5><\/h5>\n<h5><span style=\"color: #000000;\"><strong>Extrait de l\u2019article\u00a0 \u00a0\u00a0\u00a0 __________________________________________________________________________<\/strong><\/span><\/h5>\n<p><strong>Ilias Sourikopoulos 1\u00a0 \u2013\u00a0 Sara Hedayat 1\u00a0 \u2013\u00a0 Christophe Loyez 1, 2,\u00a0 \u2013\u00a0 Fran\u00e7ois Danneville 1, 2,\u00a0 \u2013\u00a0 <\/strong><strong>Virginie Hoel 1, 2,\u00a0 \u2013\u00a0 Eric Mercier 3, 4 and Alain Cappy 1, 2<br \/>\n<\/strong><\/p>\n<ul>\n<li><em><strong>1<\/strong> CNRS, Universit\u00e9 Lille, USR 3380 \u2013 IRCICA, Lille, France, <\/em><\/li>\n<li><em><strong>2<\/strong> CNRS, Universit\u00e9 Lille, ISEN, Universit\u00e9 Valenciennes, UMR 8520 \u2013 IEMN, Lille, France, <\/em><\/li>\n<li><em><strong>3<\/strong> Universit\u00e9 Grenoble Alpes, Grenoble, Grenoble, France,<\/em><\/li>\n<li><em><strong>4<\/strong> CEA, LETI, MINATEC Campus, Grenoble, France<\/em><\/li>\n<\/ul>\n<div class=\"JournalAbstract\">\n<p>As Moore\u2019s law reaches its end, traditional computing technology based on the Von Neumann architecture is facing fundamental limits. Among them is poor energy efficiency. This situation motivates the investigation of different processing information paradigms, such as the use of spiking neural networks (SNNs), which also introduce cognitive characteristics. As applications at very high scale are addressed, the energy dissipation needs to be minimized. This effort starts from the neuron cell. In this context, this paper presents the design of an original artificial neuron, in standard 65 nm CMOS technology with optimized energy efficiency. The neuron circuit response is designed as an approximation of the Morris-Lecar theoretical model. In order to implement the non-linear gating variables, which control the ionic channel currents, transistors operating in deep subthreshold are employed. Two different circuit variants describing the neuron model equations have been developed. The first one features spike characteristics, which correlate well with a biological neuron model. The second one is a simplification of the first, designed to exhibit higher spiking frequencies, targeting large scale bio-inspired information processing applications. The most important feature of the fabricated circuits is the energy efficiency of a few femtojoules per spike, which improves prior state-of-the-art by two to three orders of magnitude. This performance is achieved by minimizing two key parameters: the supply voltage and the related membrane capacitance. Meanwhile, the obtained standby power at a resting output does not exceed tens of picowatts. The two variants were sized to 200 and 35 \u03bcm<sup>2<\/sup> with the latter reaching a spiking output frequency of 26 kHz. This performance level could address various contexts, such as highly integrated neuro-processors for robotics, neuroscience or medical applications.<\/p>\n<div class=\"JournalFullText\">\n<h2>Introduction<\/h2>\n<p class=\"mb15\">Computing technology, based on binary coding, Von Neumann architecture and CMOS technology, is currently reaching certain limits (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B37\">Waldrop, 2016<\/a>). Traditional computers, the champions for the resolution of complex equation systems, have difficulties to classify\/organize data, something that the human brain seems to accomplish effectively. For this reason, research in the field of Artificial Neural Networks (ANNs) is attracting much attention and is quickly developing worldwide. At the bottom of these efforts lies the ultimate goal to realize machines that could surpass the human brain, in some aspects of cognitive intelligence. In that sense, brain research and ANNs bear the promise of a new computing paradigm.<\/p>\n<p class=\"mb15\">Currently, traditional, discrete-time, digital ANNs, fueled by the unprecedented computational capability of modern Graphics Processing Units (GPUs), represent the state-of-the-art for addressing cognitive tasks (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B29\">Oh and Jung, 2004<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B21\">LeCun et al., 2015<\/a>) such the ones encountered in computer vision applications. However, it is the more recent class of Spiking Neural Networks (SNNs), often referred to as the third generation of neural networks, that are known to be bio-realistic and more computationally potent compared with their predecessors (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B23\">Maass, 1997<\/a>). The functional similarity with the actual biological networks permits envisioning, apart from interfacing or reproducing brain processes, the implementation of circuits and systems with cognitive characteristics without explicit programming tasks. This would endow the modern generation of computers with the capacity to learn from input data.<\/p>\n<p class=\"mb15\">In SNNs, neuronal communication is carried out in the form of noise-robust, signal pulses or \u201cspikes.\u201d SNNs try to reproduce the physical characteristics of the brain, through highly connected neurons dendrites and axons. At present, two main methodologies fulfill neuro-inspired computing tasks: digital simulation and hardware implementation.<\/p>\n<p class=\"mb15\">In digital simulations, the dynamics of neuronal models are coded in software and calculated on general-purpose digital hardware. Digital simulations have the advantage that they can be reliably programmed using numerical operations of very high precision. However, their reliability comes at the cost of high circuit complexity, which is necessary for the data transfer, exchange and processing (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B8\">Cao et al., 2015<\/a>). Accordingly, the energy consumption remains still very high, especially as one juxtaposes biological data for comparison. For instance, the brain of the cat is emulated at the cost of a power dissipation in the megawatt range (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B3\">Ananthanarayanan et al., 2009<\/a>), while the animal brain actually consumes only a couple of watts.<\/p>\n<p class=\"mb15\">As far as hardware implementations of SNNs are concerned, the alternative, \u201cneuromorphic,\u201d approach consists of employing VLSI circuit technology, namely CMOS fabrication processes which can be possibly associated with more advanced device technology such as memristors (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B20\">Kim et al., 2012<\/a>). The analog hardware approach consists of a large-scale integration of silicon artificial neurons (AN) and synapses, in an attempt to produce low power neuro-inspired architectures compatible with the current electronics technology.<\/p>\n<p class=\"mb15\">The efficiency of such architectures can be revealed in contrast to the energy consumption of biological neurons (BN). Brain activity needs a continuous exchange of ions through the cell membrane and these exchanges correspond to the charge and discharge of the neuron capacitance (soma, dendrite, and axon). As a consequence, the important parameters for energy dissipation are the membrane capacitance and the voltage swing. Membrane capacitance varies considerably according to the type of neuron cell, ranging from picofarads to nanofarads for the largest ones (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B2\">Amzica and Neckelmann, 1999<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B11\">Golowasch et al., 2009<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B33\">R\u00f6ssert et al., 2009<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B35\">Tripathy and Gerkin, 2012<\/a>). Interestingly, a recent estimation of the capacitance that could be involved for computation in the human cortex is proposed in <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B12\">Hasler and Marr (2013)<\/a>. The calculations are based on a digital power model and suggest a biological system of 10<sup>12<\/sup> neurons with a 0.5 Hz average firing rate. The total capacitance value is calculated at 245 pF that is high when compared with the femtofarad order common in integrated circuits. Indeed, energy savings could be envisioned in silicon AN by aiming at reducing the capacitance and\/or the voltage swing.<\/p>\n<p class=\"mb15\">Next to a reduced capacitance, low power operation in silicon neurons can be facilitated by the physics of the MOS transistors. Indeed, as it has been observed (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B25\">Mead, 1989<\/a>, <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B26\">1990<\/a>) the nervous system uses, as its basic operation, a current that increases exponentially with the membrane voltage, similar to the current-voltage characteristic of a MOS transistor operating in subthreshold. However, the physical origins of these exponential dependencies are very different: a non-linear voltage controlled conductance in biological membrane against a current controlled by an energy barrier in the transistor. Due to this, the MOS transistor can only asymptotically approach a slope of kT\/q per e-fold of current change, while the biology is not limited as such (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B25\">Mead, 1989<\/a>, <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B26\">1990<\/a>). Even if I-V characteristics show different slopes, a bridge between the physics of biological membrane and the one of electronic devices has been established, especially when the energy and power properties are considered. This led to the advent of neuromorphic silicon neurons, which allowed neuronal spiking dynamics to be directly emulated on analog large-scale integration chips. So far, several generations of SNNs have been proposed and the reader could refer to the relevant works (<a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B27\">Misra and Saha, 2010<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B16\">Indiveri et al., 2011<\/a>; <a href=\"http:\/\/journal.frontiersin.org\/article\/10.3389\/fnins.2017.00123\/full#B12\">Hasler and Marr, 2013<\/a>) to obtain more information.<\/p>\n<p class=\"mb15\">Based upon these previous works, this paper describes the design and measurement results of a new family of silicon AN. It was designed under the guidelines of (i) a biophysically meaningful model, (ii) a minimum energy dissipation, (iii) an analog circuit that would allow a complete time variation modeling of the membrane potential and (iv) a resulting topology, that when implemented in CMOS technology, it would occupy a minimum area in order to enable large scale integration. This unique combination of characteristics resulted in a neuron topology that was measured to consume several orders of magnitude less energy than the values encountered either for BN or the AN reported so far.<\/p>\n<p class=\"mb0\">The rest of this paper is organized as follows: The \u201cMaterials and methods\u201d section will be devoted to a discussion on neuron energy efficiency, the selection of the mathematical model and the circuit topology and functionality. The circuit proposed in this paper was fabricated and characterized experimentally. Both simulation and experimental results are described in the \u201cResults\u201d section. The \u201cDiscussion\u201d section presents a comparison with the state of the art and highlights issues regarding noise, supply voltage sensitivity and temperature impact. Finally conclusions are drawn in the eponymous last section.<\/p>\n<\/div>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>Chef d\u2019\u0153uvre de l\u2019\u00e9volution, le cerveau humain est une source d\u2019inspiration pour les scientifiques. Des chercheurs de l\u2019IEMN et de l\u2019Ircica ont ainsi mis au point un neurone artificiel ultra-efficace en \u00e9nergie et qui reproduit tr\u00e8s pr\u00e9cis\u00e9ment les signaux \u00e9lectriques g\u00e9n\u00e9r\u00e9s dans le cerveau. Ces travaux sont publi\u00e9s dans la revue Frontiers in Neuroscience. Dans [&hellip;]<\/p>","protected":false},"author":2,"featured_media":23294,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[],"class_list":["post-12934","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/posts\/12934","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/comments?post=12934"}],"version-history":[{"count":0,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/posts\/12934\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/media\/23294"}],"wp:attachment":[{"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/media?parent=12934"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/categories?post=12934"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.iemn.fr\/en\/wp-json\/wp\/v2\/tags?post=12934"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}