Abstract:Abstract Spiking Neural Networks (SNNs) are gaining increasing attention for their biological plausibility and potential for improved computational efficiency. To match the high spatial-temporal dynamics in SNNs, neuromorphic chips are highly desired to execute SNNs in hardware-based neuron and synapse circuits directly. This paper presents a large-scale neuromorphic chip named Darwin3 with a novel instruction set architecture(ISA), which comprises 10 primary instructions and a few extended instructions. It supports flexible neuron model programming and local learning rule designs. The Darwin3 chip architecture is designed in a mesh of computing nodes with an innovative routing algorithm. We used a compression mechanism to represent synaptic connections, significantly reducing memory usage. The Darwin3 chip supports up to 2.35 million neurons, making it the largest of its kind in neuron scale. The experimental results showed that code density was improved up to 28.3x in Darwin3, and neuron core fan-in and fan-out were improved up to 4096x and 3072x by connection compression compared to the physical memory depth. Our Darwin3 chip also provided memory saving between 6.8X and 200.8X when mapping convolutional spiking neural networks (CSNN) onto the chip, demonstrating state-of-the-art performance in accuracy and latency compared to other neuromorphic chips.

What problem does this paper attempt to address?

This paper attempts to solve the following three main problems in order to better utilize the advantages of spiking neural networks (SNNs): 1. **Flexibility of neural models**: - Many existing brain - like chips only support a single type of neuron model, which limits their ability to simulate biological neurons and synapses. To improve flexibility, researchers have introduced some configurability to adapt to different neuron models. For example, Loihi enhances learning ability through configurable tracing and delay, and FlexLearn combines the key features of multiple models. However, these methods still have room for improvement in performance and energy efficiency. - **Solution**: Darwin3 has designed a domain - specific instruction - set architecture (ISA) that can efficiently describe multiple neuron models and learning rules, including the LIF family, Izhikevich model, and STDP, etc., and can achieve high parallelism in computational operations. 2. **Synaptic density**: - To further tap the potential of SNNs, brain - like chips need to support the representation of large - scale SNNs with more complex topologies. However, current brain - like chips pay less attention to this aspect and mainly focus on simulating the behavior of neurons and synapses. For example, TrueNorth uses a cross - bar design for synaptic connections, but is limited by a fixed fan - in / fan - out capacity; Loihi improves flexibility by axonal index - encoding the topology; Loihi2 optimizes convolutional and factorized connections, but pays less attention to other connection types. - **Solution**: Darwin3 has introduced a novel mechanism to represent the topology of SNNs, which effectively compresses the information required to describe synaptic connections, significantly reduces the overall memory usage, and supports up to 2.35 million neurons, making it the largest in its class. 3. **On - chip learning ability**: - Learning ability is an important feature of biological neural networks, but currently only a few brain - like chips support on - chip learning, and the supported learning rules are very limited. For example, BrainscaleS2 only supports a fixed learning algorithm; Loihi supports programmable rules for pre -, post -, and reward - tracing; Loihi2 extends the application range of these rules, but still cannot accommodate new learning rules. - **Solution**: Darwin3 not only supports flexible neuron - model programming and local - learning - rule design, but also provides powerful on - chip learning ability and can adapt to emerging new learning rules. In conclusion, Darwin3 aims to solve the limitations of existing brain - like chips in terms of flexibility, synaptic density, and learning ability through its domain - specific ISA, innovative topology - representation mechanism, and powerful on - chip learning ability, thereby better exploiting the advantages of SNNs.

Darwin3: A large-scale neuromorphic chip with a Novel ISA and On-Chip Learning

Darwin:a Neuromorphic Hardware Co-Processor Based on Spiking Neural Networks

DaDianNao: A Machine-Learning Supercomputer

A 64K-Neuron 64M-1b-Synapse 2.64pJ/SOP Neuromorphic Chip With All Memory on Chip for Spike-Based Models in 65nm CMOS

Multicore Spiking Neuromorphic Chip in 180-nm with ReRAM Synapses and Digital Neurons

Mapping Very Large Scale Spiking Neuron Network to Neuromorphic Hardware.

Scalable NoC-based Neuromorphic Hardware Learning and Inference

An Efficient Neuromorphic Implementation of Temporal Coding-Based On-Chip STDP Learning

MorphBungee: A 65-nm 7.2-mm2 27-μJ/image Digital Edge Neuromorphic Chip with On-Chip 802-frame/s Multi-Layer Spiking Neural Network Learning

Toward Robust Cognitive 3D Brain-Inspired Cross-Paradigm System

TripleBrain: An Edge Neuromorphic Architecture for High-accuracy Single-layer Spiking Neural Network with On-chip Self-organizing and Reinforcement Learning

ActiveN: A Scalable and Flexibly-Programmable Event-Driven Neuromorphic Processor

A 28-Nm 0.34-Pj/sop Spike-Based Neuromorphic Processor for Efficient Artificial Neural Network Implementations

A Novel Neuromorphic Hardware Using Area-Efficient Chain RRAM-Based Synapses and Compact Neurons with (Anti-) Integration Scheme

Brain-Inspired Hardware for Artificial Intelligence: Accelerated Learning in a Physical-Model Spiking Neural Network

Neuromorphic chip integrated with a large-scale integration circuit and amorphous-metal-oxide semiconductor thin-film synapse devices

Design Space Exploration of Hardware Spiking Neurons for Embedded Artificial Intelligence

eBrainII: a 3 kW Realtime Custom 3D DRAM Integrated ASIC Implementation of a Biologically Plausible Model of a Human Scale Cortex

Scalable Digital Neuromorphic Architecture for Large-Scale Biophysically Meaningful Neural Network With Multi-Compartment Neurons

Energy-efficient SNN Architecture using 3nm FinFET Multiport SRAM-based CIM with Online Learning

A Neuromorphic Spiking Neural Network Using Time-to-First-Spike Coding Scheme and Analog Computing in Low-Leakage 8T SRAM