Chinese Academy of Sciences Unveils New Neuromorphic Computing System
On June 1, TapTechNews reported that the human brain can run a very complex and huge neural network, but the power consumption is only 20 watts, which is much less than that of existing AI systems. Therefore, in today's era of accelerating computing power competition and rising energy consumption, it has become a highly potential direction to develop a new type of intelligent computing system by referring to the low-power consumption characteristics of the human brain.
The Institute of Automation, Chinese Academy of Sciences announced that the research group of Li Guoqi and Xu Bo proposed the concept of 'neuromorphic dynamic computing' and cooperated with domestic scientific and technological enterprises and other units to design a neuromorphic SoC system named Speck that can realize dynamic computing with an algorithm-software-hardware collaborative design, so as to realize dynamic computing based on the attention mechanism, which can achieve 'no input, no power consumption' at the hardware level and 'when there is input, dynamically adjust the calculation according to the importance degree of the input' at the algorithm level. The power consumption in typical visual scene tasks can be as low as 0.7 milliwatts, further exploiting the potential of neuromorphic computing in performance and energy efficiency.
Related research results have been published online in Nature Communications (TapTechNews attached DOI: 10.1038/s41467-024-47811-6).
Speck is an asynchronous sensing-computing-integrated neuromorphic SoC, adopting a fully asynchronous design. It integrates a dynamic visual sensor (DVS camera) and a neuromorphic chip on one chip, with extremely low quiescent power consumption (only 0.42 milliwatts).
Speck can sense visual information with a microsecond-level time resolution, and the fully asynchronous design abandons the global clock control signal, avoiding the energy consumption overhead caused by clock glitching and only triggering sparse addition operations when there is an event input.
The official said that the practice of this work confirms that the fusion of high and low abstract level brain mechanisms can further stimulate the potential of brain-like computing, providing positive inspiration for the future integration of various advanced neural mechanisms generated in the brain evolution process into neuromorphic computing.
