It sounds pretty cool - Chinese scientists have developed a low-power brain-like chip
[Copy link]
A piece of news I saw today: Chinese scientists have developed a low-power brain-like chip. It sounds amazing. What do you think about it?
The following is the content of the news:
Source: The content is compiled by Semiconductor Industry Observer (ID: icbank) from China News Service and others, thank you.
According to the Institute of Automation of the Chinese Academy of Sciences, a research team of the institute has collaborated with other units to design a new brain-like neuromorphic system-on-chip, Speck, which demonstrates the natural advantages of neuromorphic computing when integrating high-level abstract brain mechanisms. The related research was recently published online in the international academic journal Nature Communications.
"The human brain is a very complex and large neural network system with a total power consumption of only 20 watts, which is far less than existing artificial intelligence systems." Li Guoqi, a researcher at the Institute of Automation, Chinese Academy of Sciences, said that therefore, at a time when the competition in computing power is accelerating and energy consumption is rising, learning from the low-power characteristics of the human brain to develop new intelligent computing systems has become a very promising direction.
An important function of the human brain is to dynamically allocate its limited attention resources according to the importance of external stimuli. Important stimuli tend to receive more attention, which is called the attention mechanism. This study proposed the concept of "neuromorphic dynamic computing", applying the high-level attention mechanism in the human brain to the design of brain-like chips, further exploring the potential of brain-like chips in performance and energy efficiency.
Li Guoqi said that Speck integrates dynamic visual sensors and neuromorphic chips on a single chip, with extremely low resting power consumption. The power consumption of typical visual scene tasks can be as low as 0.7 milliwatts, providing a brain-like intelligent solution with high energy efficiency, low latency and low power consumption for artificial intelligence applications.
In this study, the collaborative team proposed the concept of "neuromorphic dynamic computing" by designing a brain-like neuromorphic chip, Speck, to achieve dynamic computing based on the attention mechanism. At the hardware level, it achieves "no input, no power consumption", and at the algorithm level, it achieves "when there is input, the calculation is dynamically adjusted according to the importance of the input". In typical visual scene tasks, the power consumption can be as low as 0.7 milliwatts, further exploring the potential of neuromorphic computing in performance and energy efficiency.
Speck is an asynchronous sensing and computing brain-like neuromorphic system-level chip. It adopts a fully asynchronous design and integrates a dynamic visual sensor (DVS camera) and a brain-like neuromorphic chip on a single chip. It has extremely low resting power consumption (only 0.42 milliwatts). It can perceive visual information with a time resolution of microseconds. It is designed in a fully asynchronous way to abandon the global clock control signal, avoid the energy consumption caused by clock flips, and trigger sparse addition operations only when there is an event input.
Li Guoqi pointed out that in response to the "dynamic imbalance" problem of spiking neural networks (SNNs) at a higher level, such as the inability to adjust the pulse emission according to the difficulty of the input in the time dimension, this study uses a neuromorphic spiking dynamic computing framework based on the attention mechanism to achieve differentiated dynamic responses to different inputs at multiple granularities. At the same time, the Speck software tool chain programming framework supports the training and deployment of dynamic computing spiking neural network algorithms.
The experimental results of this study show that the attention mechanism can enable the spiking neural network to have dynamic computing capabilities, that is, to adjust its pulse emission mode according to the difficulty of the input to solve the "dynamic imbalance" problem, significantly reducing power consumption while improving task performance. On a dynamic visual sensing dataset, Speck, which integrates pulse dynamic computing, improved task accuracy by 9% while reducing average power consumption from 9.5 milliwatts to 3.8 milliwatts.
|
提升卡
变色卡
千斤顶
京公网安备 11010802033920号
