Significant energy savings through nervous systems

Study from Graz University of Technology and Intel
Significant energy savings through nervous systems

In a recent study, Graz University of Technology and Intel demonstrated that using neural shape technology in large deep learning networks leads to significant energy savings.

The Institute for Fundamentals of Information Processing at TU Graz, together with Intel Labs, was able to experimentally prove that a large neural network can process sequences on neural machines four to sixteen times faster than on conventional machines. The results are based on Intel’s first generation Loihi research chip. Findings from neuroscience are used to build chips based on the human brain model. The results of the research were published in the journal Nature Machine Intelligence. published.

Focus on energy consumption

The massive consumption of energy is a major obstacle to a broader application of intelligent AI methods. Neuromorph technology can provide a nudge in the right direction here: it relies on the human brain, which has impressive values ​​in terms of energy efficiency. Roughly a hundred billion neurons only need about 20 watts to process information – hardly more power than the average energy-saving lamp.

A team from Graz University of Technology and Intel Laboratories focused their work on algorithms that work with temporal processes. For example, the system had to answer questions about a previously told story and pick up relationships between objects or people from context. 32 Loihi chips were used as devices.

Biology as a resource for ideas

“Our system is four to sixteen times more energy efficient than other AI models on conventional machines,” explains Philip Blank, a doctoral student at the Graz University of Technology. With the next generation of Loihi devices, the efficiency can increase even more.

“Our work with the Graz University of Technology provides further evidence that neuromorphic technology can improve the power efficiency of today’s deep learning workloads by rethinking its implementation from a biology perspective,” adds Mike Davies, Director of the Intel Neuromorphic Computing Lab.

(No.: 48393330)