To address the shortcomings of traditional thermal management algorithms, such as insufficient accuracy under dynamic loads, poor generalization of DRL algorithms with few shots, and superficiality in the fusion of GNN and DRL, this paper proposes a graph attention-enhanced meta-learning deep reinforcement learning algorithm (GAM-SSDRL). The algorithm constructs a three-level closed-loop architecture of “thermal field graph modeling - few-shot decision-making dynamic collaborative optimization.” It accurately captures spatial correlations of the thermal field through a graph attention network, achieves efficient few-shot training based on an improved MAML framework, and adaptively matches feature extraction and decision execution through a dynamic collaborative strategy. Experiments based on a 16 -core CPU chip simulation model show that GAM-SSDRL performs optimally under high, medium, and low load scenarios, with an average temperature deviation as low as 1.6° C, representing reductions of 50.0 %, 23.8 %, and 15. 8 % compared to traditional PID, conventional DRL, and GNN+DRL, respectively. The thermal runaway response time is 44.3 ms, a reduction of over 50.6 %; the average chip power consumption is 78.9 W, a reduction of 18.3 %; and only 100 training samples are needed to achieve the target accuracy, improving sample efficiency by over 54.5 %. This algorithm fills the technical gap in precise control of dynamic thermal fields with small sample sizes, providing a new solution for thermal management of high-power-density chips, and has significant theoretical and engineering value.