Artificial intelligence (AI) has seen explosive growth in recent years, but despite major progress, the power required to run AI algorithms continues to increase.
In stark contrast to this, the human brain only requires around 20W to perform more than 10 quadrillions (10,000,000,000,000,000) operations. This is 12 orders of magnitude better than modern supercomputer technologies.
“Although modern CPUs and GPUs can perform massive volumes of complex arithmetic-logic calculations, the human brain is still far superior when it comes to cognitive applications. That’s why we’re conducting intensive research into developing new hardware that mimics some functions of the human brain, with neurons, synapses and neural networks, known as brain-inspired computing (BIC),” says Assistant Professor Hooman Farkhani, an expert in AI hardware at the Department of Electrical and Computer Engineering at Aarhus University.
He continues:
“But even though we’ve managed to drastically reduce the energy consumption of hardware running AI algorithms, there’s still a long way to go before BICs are as efficient as the human brain when it comes to size and energy efficiency.”
Hooman Farkhani has just received a grant of DKK 1.9 million from the Villum Experiment programme under the Danish Villum Foundation for his new project Spin-Grain, which is looking into the development of a nano-sized BIC system.
“I believe, I can take brain-inspired computing a big step forward by scaling down the AI into a single nano-device for the first time. If we succeed, we’ll have the first BIC system that is no larger than a grain of dust and with energy consumption that is so small that energy can be harvested directly from the surrounding environment. In other words, no power supply will be needed, and this will pave the way for a range of new, previously impossible AI applications,” he says.
Hooman Farkhani is part of the Aarhus University ICE-Lab (Integrated Circuits and Electronics Laboratory) where his focus is on developing and designing emerging computer architecture suitable for AI algorithms.