Design of a Highly Flexible Hybrid Neural Network Inference Platform with 10 Million Neurons

A large-scale neural network platform that explores brain functions while deploying artificial networks has become an urgent research objective. However, large-scale platforms struggle to balance versatility, flexibility, and performance. This paper proposes a hybrid network platform comprising 1280 cores, supporting the deployment of 10 million neurons and 80 billion synapses. Functionally, the cores utilize a hybrid neuron model, enabling neuromorphic and traditional artificial intelligence computation and supporting the mapping of many structures, especially the Transformer. In terms of performance, the proposed architecture features a bandwidth-matching design, employing near-memory computing and multi-channel high-speed interfaces to alleviate bottlenecks, achieving a 9.0 times speed improvement compared to off-chip computation. For flexibility, the architecture uses flexible core cluster designs, incorporating variable-scale deployment and intra-cluster fusion, providing up to a 7.1 times increase in computational utilization and a maximum 341 times reduction in latency for the same network deployment. The architecture successfully deployed multiple networks for recognition on an emulator. This architecture supports flexible hybrid inference, providing comprehensive hardware conditions for deploying large-scale neural networks.