A research team at Bernal Institute in the 51±¾É«, have designed a new molecular structure that hold the potential to revolutionise artificial intelligence (AI) hardware by drastically improving computational speed and energy efficiency.
The research, which was recently published in the leading scientific journal Nature, was conducted as part of an international collaboration involving scientists from around the world including the Indian Institute of Science (IISc) and Texas A&M University.
Led by Professor Damien Thompson, director of SSPC, the Research Ireland Centre for Pharmaceuticals, the team discovered new techniques to probe, control and tailor materials at a molecular level.
The breakthrough could transform the future of computing.
The discovery has culminated in the development of a cutting-edge hardware platform that achieves previously unrealised improvements in speed and energy efficiency for AI applications. The platform, inspired by the human brain's processes, marks significant progress for neuromorphic computing, a technology that mimics the neural networks of the brain.
“The design draws inspiration from the human brain, using the natural wiggling and jiggling of atoms to process and store information," said Professor Thompson.
"As the molecules move around their crystal lattice, they create a multitude of individual memory states. This molecular movement can be mapped to unique electrical states, which can be read and written like in conventional computers, but with massively improved energy and space efficiency."
Professor Thompson said he sees a future where computing is easily integrated into everyday objects: “The ultimate aim is to replace what we now think of as computers with high-performance ‘everyware’—energy-efficient and eco-friendly materials providing distributed, ubiquitous information processing throughout the environment, from clothing to food packaging and building materials.â€
The breakthrough could have wide-reaching benefits, including reducing the energy consumption of data centres, improving memory-heavy applications like digital maps and online gaming, and advancing AI technologies.
Until now, neuromorphic platforms have been limited to low-accuracy operations, such as inferencing in AI neural networks. However, UL’s new discovery overcomes this limitation, achieving high resolution and performing resource-intensive tasks with unprecedented energy efficiency—reaching 4.1 tera-operations per second per watt (TOPS/W).
Professor Sreetosh Goswami of IISc, the project lead, said: “By precisely controlling the vast array of available molecular kinetic states, we created the most accurate, 14-bit, fully functional neuromorphic accelerator integrated into a circuit board that can handle signal processing, AI and machine learning workloads such as artificial neural networks, auto-encoders, and generative adversarial networks.
“Most significantly, leveraging the high precision of the accelerators, we can train neural networks on the edge, addressing one of the most pressing challenges in AI hardwareâ€.