The latest development in neuromorphic computing research focuses on the Spiking Neural Network (SNN), offering a unique method for simulating brain functions. Unlike traditional neural networks, SNNs operate on discrete, event-driven signals that closely mimic biological processes.
To harness the benefits of SNNs, specialized neuromorphic computing chips are being designed to revolutionize traditional computing architecture. These chips aim to address storage and power limitations in the post-Moore era, offering a promising solution for advanced computational tasks.
Despite the potential of SNNs, researchers encounter challenges in modeling diverse behaviors accurately, scaling synaptic connections effectively, and enabling on-chip learning capabilities. To tackle these obstacles, Professor Gang Pan’s team at Zhejiang University and Zhejiang Lab collaborated to develop the Darwin 3 neuromorphic chip, part of the Darwin series.
Their research, published in the National Science Review, introduces a new instruction set architecture (ISA) tailored for neuromorphic computing. This ISA facilitates rapid state updates and parameter loading, enabling efficient model construction and learning rules implementation.
Moreover, the team devised an innovative connection mechanism enhancing on-chip storage efficiency, supporting over 2 million neurons and 100 million synapses on a single chip. These advancements lay a solid hardware foundation for constructing brain-scaled neural networks, a critical aspect of mimicking human brain connectivity.
Further advancements include on-chip learning capabilities for Darwin3, enabling efficient handling of new information and dynamic environments while running spiking neural networks. Recent experiments showcase Darwin 3’s exceptional ability to support various SNNs and adaptability in complex scenarios, setting it apart from other neuromorphic chips.
The development of Darwin 3 signals a significant milestone in neuromorphic computing, promising to elevate artificial intelligence capabilities to new heights. This groundbreaking research underscores the importance of innovation in neuromorphic computing technology, paving the way for future advancements in artificial intelligence and computational neuroscience.
