Darwin3: a large-scale neuromorphic chip with a novel ISA and on-chip learning

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) offer biological plausibility and computational efficiency.
• Neuromorphic chips are crucial for executing SNNs in hardware, mimicking neural dynamics.

Purpose of the Study:

• To present Darwin3, a large-scale neuromorphic chip designed for advanced SNN applications.
• To introduce a novel instruction set architecture and routing algorithm for efficient SNN processing.

Main Methods:

• Designed Darwin3 with a mesh architecture and a novel instruction set.
• Implemented a synaptic connection compression mechanism to reduce memory footprint.
• Utilized flexible neuron model programming and local learning rule designs.

Main Results:

• Darwin3 supports up to 2.35 million neurons, the largest scale reported.
• Achieved code density improvements up to 28.3× and significant fan-in/fan-out enhancements.
• Demonstrated memory savings of 6.8× to 200.8× for convolutional SNNs.
• Exhibited state-of-the-art accuracy and latency compared to existing neuromorphic chips.

Conclusions:

• Darwin3 represents a significant advancement in large-scale neuromorphic computing for SNNs.
• The chip's architecture and compression techniques enable unprecedented neuron scales and efficiency.
• Darwin3 offers a powerful platform for SNN research and applications, outperforming current alternatives.