Neuromorphic chips are more energy efficient for deep learning


The study was conducted by the Institute of Theoretical Computer Science of the Graz University of Technology (TU Graz) in Austria using the Intel Loihi 2 neuromorphic chip, which contains about a million artificial neurons. According to the study, Intel’s neuromorphic chips are 16 times more energy efficient in deep learning tasks than existing hardware. The test was conducted using 32 Intel Loihi 2 chips and led to an important discovery in the field of artificial intelligence.

According to the researchers , “The astounding achievements of DeepMind’s high-tech AlphaGo and AlphaZero AI systems require thousands of parallel processors, each of which can consume about 200 watts”.

“Our system is 4-16 times more energy efficient than other AI models on conventional equipment. Further efficiency improvements will occur thanks to the next generation of Loihi processor”Philipp Planck, a graduate student at the Institute of Theoretical Computer Science at the Technical University of Graz, said.

The experts worked with algorithms involving temporary processes. One of the examples given was a system that answered questions about a previously told story or determined relationships between objects or people from context. The model imitated a person’s short-term memory. In the study, experts linked two types of deep learning networks:

  • feedback neural networks that are responsible for short-term memory and can control objects, recognize text, speech and describe images;
  • direct-link networks capable of predicting, recognizing images and classifying objects.

According to Intel, neuromorphic chip technology can be integrated into the CPU to add energy-efficient AI processing to the system. In addition, access to neuromorphic processors can be provided in the form of a cloud service.

Start a discussion …