An international research team guided by the Swinburne University of Technology has developed the world’s most powerful neuromorphic processor for AI. This is a major advance in the field of artificial intelligence (AI). It runs at an astonishing speed of more than 10 trillion operations per second (TeraOps / s), which means it can handle very large data.
The work was published in the journal Nature.
Under the direction of Professor David Moss Swinburne, Dr. Xingyuan Xu, and Dear RMIT University Professor Arnan Mitchell, the team accelerates computational speed and processing power. They are capable of creating an optical neuromorphic processor that can run more than 1,000 times faster than the previous one. The system can also process images at a very large scale, which is important for facial recognition because previous optical processors failed in this regard. Professor Moss is director of the Swinburne Center for Optical Science and has been recognized by Australia as one of Australia’s best physicists and mathematics researchers in the field of optics and photonics.
“This breakthrough is being made with an optical micro comb, like our world record for internet speed, which was reported in May 2020,” he said.
Other top processors and micro-combs
The best electronic processors like Google TPU can run over 100 TeraOps / s. However, it required tens of thousands of parallel processors while the team’s optical system was based on only one processor. This is achieved through the use of a new technique in which data are simultaneously interwoven in time, wavelength, and spatial dimensions via an integrated micro-comb source.
For those unfamiliar with micro-combs, this is a new device consisting of hundreds of high-quality infrared lasers on a single chip. Compared to other optical sources, micro combs are much faster, lighter, and less expensive.
“In the 10 years since I discovered them, integrated micro-silver chips have become very important and it is very exciting to see them enabling major advances in communication and information processing,” said Professor Moss. “Micro-combs offer us great promise to meet the world’s insatiable information needs.”
Dr. Sue co-authored the study and is a Swinburne Fellow and Postdoctoral Fellow in the Department of Electrical Engineering and Computer Systems at Monash University.
“This processor can serve as a universal ultra-high bandwidth front-end for any neuromorphic hardware – optical or electronic – and enables machine learning with real-time, very high bandwidth data within range,” he said. Dr. Xu.
“Right now we are looking at what processors will look like in the future. This shows us how dramatically we can measure the efficiency of our processors through the use of innovative micro combs,” he continued.
RMIT Professor Mitchell explains, “This technology applies to all forms of processing and communication – it will have a big impact. In the long term, we hope to implement a fully integrated system on a single chip, which will significantly reduce costs and energy consumption. “
Professor Damien Hicks supports the research team and comes from Swinburne and the Walter and Elizabeth Hall Institute.
“Convolutional neural networks are critical to the artificial intelligence revolution, but existing silicon technology is increasingly a bottleneck in processing speed and energy efficiency,” said Professor Hicks.
“This breakthrough demonstrates how new optical technologies make such networks faster and more efficient and is an in-depth demonstration of the benefits of interdisciplinary thinking because it gives you inspiration and the courage to take ideas from a field and use them to solve fundamental problems using a single problem. on another matter, “he continued.