Our ever-growing sea of connectivity powered by technologies such as Big Data, AI, 5G, and IoT has created an enormous tsunami of data that is, unfortunately, playing a significant role in killing our planet. A group of scientists has warned that the rapidly expanding Internet and our mammoth data consumption could be responsible for a staggering 20% of global electricity consumption and 5.5% of all CO2 emissions by 2025.
It’s an energy black hole that could quickly put paid plans to limit global warming by 1.50C above pre-industrial levels by 2100 in jeopardy.
That’s why the news that researchers have unveiled a technology that can potentially lower AI’s massive carbon footprint by up to 30x is bound to bring a smile to the faces of environmentalists everywhere.
Scienmag has reported that scientists at the Cockrell School of Engineering at The University of Texas have discovered that ditching silicon for magnetic circuits in AI neural networks could make smart computers much more energy efficient.
Smart Neural Networks
The researchers have discovered that by using magnetic components instead of silicon in neural training networks (software or hardware systems that operate like the neurons of the human brain), energy consumption could drop by a factor of 20x-30x.
Jean Anne Incorvia, an assistant professor in the Cockrell School’s Department, and second-year graduate student Can Cui have discovered that carefully spacing magnetic nanowires to act as artificial neurons invokes a process known as lateral inhibition.
This is a natural process used by human neurons, whereby the most activated neurons win out and suppress slower firing ones. Achieving lateral inhibition using current silicon systems requires extra layers of circuitry within computers, which can significantly increase costs and space requirements as well as vastly multiply energy consumption.
Their discovery could have noteworthy ramifications in our quest to control climate change.
Building Artificial Intelligence, or AI, systems usually involves deep learning; a type of machine learning method that is based on artificial neural networks.
The process of deep learning takes place in two steps: training and inference. These are extremely computational-intensive processes, such that AI was once thought to belong to the realm of science fiction– until superfast GPUs and FGPAs with matching parallel computing abilities eventually caught up.
Nowadays, AI plays more quotidian roles in our lives, such as Google’s RankBrain system that uses AI to better understand search queries; programmatic ads, malware detection, spam filtering, object recognition, speech recognition, natural language processing (NLP), translation and also in those ubiquitous (and annoying) chatbots among other tasks. But it’s easy to forget that AI has come a long way–and has become far dirtier–since the days when IBM’s Deep Blue supercomputer beat the world’s then best chess player Garry Kasparov in a series of chess matches that the machine won 4–2.
Indeed, Data published by OpenAI shows that the computing power by AI systems that performed more recent landmarks such as defeating humans at Go has doubled roughly every 3.4 months, increasing exponentially by 300,000-fold from 2012 to 2018.
Putting it mildly, AI systems and silicon neural training networks are an energy hog of absurd proportions.
A paper by MIT researchers titled “Energy and Policy Considerations for Deep Learning in NLP” says training AI models can emit more than 626,000 pounds of CO2 equivalent–or nearly five times the lifetime emissions of the average American car, including the manufacture of the car itself.
The AI market is one of the fastest-growing tech sectors, with the AI software market expected to expand a sizzling 154%Y/Y to $22.6 billion in 2020 and maintain triple-digit growth over the next five years at the very least. It’s this kind of exponential growth that has prompted academics to challenge the notion that we can considerably reduce carbon emissions by simply cutting down on waste and increasing the efficiency of our IT systems.
But the kinds of energy savings that the University of Texas researchers have demonstrated are transcendental, not merely incremental, thus offering hope that our climate goals can still be achieved.
Depending on how fast the University of Texas researchers are able to commercialize their magnetic neural networks technology, we could be able to turn back the hands of time nearly a decade by dramatically lowering the carbon footprint of modern AI systems.
Meanwhile, other promising energy-saving technologies such as MIT’s Photonic Chips that use light instead of electricity could also help in the race to combat climate change.
And when the economy opens back up in the wake of the global COVID-19 pandemic, combating climate change is likely to emerge as an even greater goal, with the impetus behind it than only a pandemic can bring.
By Alex Kimani for Oilprice.com
More Top Reads From Oilprice.com: