This post is also available in: heעברית (Hebrew)

Artificial Neural Networks (ANNs) are a type of information processing technique based on emulating processing systems of biological brains. It has been applied in fields ranging from pattern recognition and signal processing to decision support systems and autonomous control. Spiking Neural Networks (SNNs) are a subset of ANNs that can perform information processing based in discrete time spikes. It’s closer to a biological brain the classic ANNs, and has the potential to achieve much better performance to power ratios.

Researchers from Zhejiang University and Hangzhou Dianzi University have developed a neuromorphic hardware coprocessor based on SNNs called the Darwin Neural Processing Unit (NPU). It’s based on standard CMOS (complementary metal-oxide semiconductor) technology.

Intelligent devices are becoming ubiquitous, especially with the rise of the Internet of Things. They provide many benefits, but also raise many challenges in running complex algorithms on small devices.

The Darwin NPU aims to address these challenges. It will provide hardware acceleration to complex intelligent algorithms, with the intended application in small, resource constrained, low-power embedded devices. It’s fabricated in a standard 180nm CMOS process, supporting a maximum of 2048 neurons, more than four million synapses, 15 possible synaptic delays, and is highly configurable.

Darwin provides a real-world demonstration of the feasibility of real-time execution of SNN in resource constrained embedded systems. Potential applications include intelligent hardware systems, and robots. As the Darwin chip uses spikes for information processing and transmission, much the same as biological brains, it may even be suitable for the analysis of biological spiking neural signals. This could allow for the creation of interfacing to connect with animal or even human brains.