As the world becomes increasingly digitalized, data centers and data transmission networks are emerging as an important source of energy demand, each accounting for about 1% of global electricity use. Global data center electricity demand in 2019 was ∼200 TWh, or around 0.8% of global final electricity demand.
So far, strong growth in demand for data center services continues to be offset by ongoing efficiency improvements for servers, storage devices, network switches and data center infrastructure, as well as a shift to much greater shares of cloud and hyperscale data centers.
The emphasis here is on “so far” – the big question is how long this offset will work and when, not if, data traffic growth will exceed efficiency gains. Consider that global internet traffic surged by almost 40% in the short period between February and mid-April 2020, driven by growth in video streaming, video conferencing, online gaming, and social networking.Demand for data and digital services is expected to continue its exponential growth over the coming years, with global internet traffic expected to double by 2022 to 4.2 zettabytes per year (4.2 trillion gigabytes). Between 2019 and 2022, traffic from internet video is projected to more than double to 2.9 zettabytes, while online gaming is projected to quadruple to 180 exabytes (180 billion gigabytes). Together, these streaming services are projected to account for 87% of consumer internet traffic in 2022.
Emerging digital technologies such as machine learning (AI), blockchain (it is estimated that Bitcoin miners consumed 50-70 TWh in 2019, or 0.2-0.3% of global electricity use), 5G, and virtual reality are also poised to raise demand for data services.
In addition, the number of mobile internet users is projected to increase from 3.8 billion in 2019 to 5 billion by 2025, while the number of Internet of Things (IoT) connections is expected to double from 12 billion to 25 billion in that period. These trends are driving exponential growth in demand for data center and network services.
“The exponential deluge of data has only just begun in earnest, Moore’s scaling is tapering out, and the efficiency of transistors has not improved in the past few generations, setting the stage for a watershed moment in computing, and by direct extrapolation, energy consumption and the global information economy,” says Dr Sreetosh Goswami, a Research Fellow at National University of Singapore’s Nanoscience & Nanotechnology Initiative (NUSNNI).
“There is still a silver lining to this,” he adds. “Right from its early days, the semiconductor industry has primarily bought into one dominant model for computers—the von Neumann architecture. This model adopted by almost all the major industries has subsequently forced the market to adapt to a fixed technology rather than creating opportunities for supporting fit-for-purpose technology that the market needs. This challenge faced by the computing industry today could enable a new era scientific innovations and creativity, bringing different disciplines together like never before in mainstream electronics.”
Today, most computer scientists believe that the brain is a hallmark of millions of years of evolution that vastly outperforms any existing man-made computing platform. For developing cognition in hardware platforms through neuromorphic computing, drawing inspiration from brain designs could be the most efficient yet the most daunting route.
The biggest challenge for the development of brain-inspired computing is that the brain is too complex to simply emulate on a chip. The human brain can be seen as a nonlinear dynamic system that acquires large amounts of data through multiple input channels, performing filtering functions, transmitting data through a densely interconnected network, storing key information in both short- and long-term memories, learning by dynamically correlating both incoming and stored information, and making decisions in a dynamically evolving and uncertain environment, while consuming very little power. This massive interconnectivity, redundancy, local activity, remarkable logic complexity, and functional non-linearity makes the brain a computing masterpiece.
Many researchers in this field see memristors as a key device component for neuromorphic computing. Memristor – or memory resistor – devices are non-volatile electronic memory devices that were first theorized by Leon Chua in the 1970’s. However, it was some thirty years later that the first practical device was fabricated in 2008 by a group led by Stanley Williams at HP Research Labs.
Following in the footsteps of this early work, Goswami and his fellow scientists, led by Professor T. Venkatesan, then the director of NUS Nanoscience and Nanotechnology Institute (NUSNNI), are working on a non-conventional genre of memristors made of metal-organic molecular complexes that shows great promise for ultralow energy, high density computing platforms. Building low-energy computing platforms beyond Moore was an initiative Professor Venkatesan took up in 2012. After starting the journey with ferroelectric tunnel junctions and oxide memristors, he got interested in the molecular memristor devices after Goswami joined his group in 2015 as a graduate student. This team is now closely working with Professor Stanley Williams, who is currently the HPE Chair Professor at Texas A&M University, to explore technological potential of this molecular system.
Fundamentally, these devices could be a million times more power efficient than any commercial device that exists today.
“The switching energy of our devices is at least 3 orders of magnitude lower than the best examples in oxide memristors,” he tells Nanowerk. “Additionally, our molecular redox mechanisms are very different from those in conventional electronic devices, and we have evidence that our molecular mechanisms might enable dynamic responses optimal for neuron-like computation. Our systems can offer a credible route to enable brain inspired functionalities on a chip.”
Major challenges with conventional molecular memristors have been that their performances have been poor, and characterizations – if performed at all – have been inconsistent with proposed molecular mechanisms. Using memristors based on metal-azo complexes, the NUSNNI team was able to overcome both these limitations.
In a paper in Nature Materials in 2017 (“Robust resistive memory devices using solution-processable metal-coordinated azo aromatics”), they established that molecular mechanisms can be reproducible and consistent over hundreds of devices. In this work they also developed an in situ spectroscopic platform that enabled them to measure molecular mechanisms on the fly in a device in operation.
Using this in operando spectroscopic platform, the researchers discovered a phenomenon that had been an elusive goal in condensed matter physics for over 50 years. Reporting earlier this year in Nature Nanotechnology (“Charge disproportionate molecular redox for discrete memristive and memcapacitive switching”), they demonstrated a nanoscale device based on a unique material platform that can achieve optimal digital in-memory computing, while being extremely energy efficient.
“We show that we can drive valence symmetry breaking in our molecular film at ambient conditions just using applied voltage,” Goswami says. “To date, all other examples of valence symmetry breaking have been shown at low temperature or high pressure, which restrict their potential applicability. By enabling this phenomenon in ambient conditions, we facilitated the co-existence of a ternary memristor and a binary memcapacitor in the same nanoscale device.”
This provides the first condensed matter route to produce memcapacitance and its co-existence with memristance enables multiple state transitions, which is one of the highly sought functionalities by circuit theorists.
“Our studies are basically bridging condensed matter physics with elemental circuit properties,” Goswami notes.
The class of robust metal-organic molecular system this team is using is a brainchild of Professor Sreebrata Goswami from the Indian Association for the Cultivation of Science (IACS), Kolkata. These molecules consist of three key components:
A transition metal center that offers structural and valence stability to the molecular system;
a set of ligands where charge can be transferred to the azo group, which has three natural charge states (redox states); and
one or two counter ions that spatially occupy specific regions depending upon the redox states of the ligands offering the system hysteresis.
Since the redox states determine the electron transport, the system exhibits different conductance states depending on the redox states of the ligands, which are voltage dependent. Professor Damien Thompson from the University of Limerick in Ireland has been working with the team to theoretically model these molecular devices.
“One really important factor amidst everything is that we have a Raman spectroscopic tool that can be used to precisely monitor various redox states,” Goswami points out. “This provides a great platform for building our understanding of the device.”
While talking presenting these molecular mechanisms to the research community, one inevitable question the NUSNNI team was asked in every other conference and by almost every referee was whether their switching was spatially uniform.
This is one of the natural consequences of a molecular mechanism but remains a puzzle because all previous reports attempting to characterize a nanoscopic picture of switching in molecular films and even oxide-based systems show random current spikes, contrary to one’s expectation. This means either the characterization is flawed, or the mechanism is non-molecular.
In their recent report in Advanced Materials (“Nanometer-Scale Uniform Conductance Switching in Molecular Memristors”), the researchers resolve this long-standing conundrum by demonstrating 100% spatially homogeneous current switching with <7nm resolution.
"This settles a long-standing debate on the authenticity of organic memristive switching," says Goswami. "We are really glad that this paper received excellent complements from the referees and also got selected as a cover page of Advanced Materials."
Based on the knowledge developed in their current studies, the researchers are now working on closing the gap between device development and eventual implementation of a functional, brain-inspired computing architecture to harness the capabilities and exclusivity this material system can offer.
According to the researchers, right now the biggest challenge is scaling up to large-area circuits involving multiple devices. Already, they are making progress in this direction by adopting ideas from existing literature of flexible electronic devices and OLED (Applied Materials Today, "Colossal current and voltage tunability in an organic memristor via electrode engineering"). They are also collaborating with several academic and industrial groups.
"While emulating all functions of a brain is way beyond the reach of any existing technology at the moment, we are taking promising baby steps towards that goal," Goswami concludes. "Our multiple conditional switching transitions are coming out to be handy in emulating several brain functions that are unprecedented using other platforms. New algorithms and computing platforms will have to be developed to fully exploit our technology, which will present exciting research opportunities. The future looks extremely bright!"