关键词: 光网算用一体化, 大模型调度, 算网协同, 全链仿真, 广域智算

Abstract:

With the rapid iteration of Artificial Intelligence （AI） large models and the evolution of wide-area intelligent computing networks，the traditional single computing-optical synergy architecture can no longer meet the stringent requirements of large model training and inference for full-link resource coordination across the “optical-network-computing-application” continuum. Focusing on this core challenge，this paper conducts research on the key technologies and implementation approaches of optical-augmented computing，and designs an integrated optical-network-computing-application scheduling platform scheme based on Optical Circuit Switching （OCS）. A five-layer integrated architecture consisting of the adaptation layer，perception layer，control layer，service layer，and full-chain simulation layer is constructed，forming a closed-loop scheduling mechanism of “perception-decision-execution”. This scheme can effectively support the differentiated demands of diverse AI services and facilitate the construction of an efficient and reliable wide-area intelligent computing network ecosystem.

Key words: optical-network-computing-application integration, large model scheduling, computing-network collaboration, full-chain simulation, wide-area intelligent computing