<p>The ever-increasing integration density and high operating frequency of 3D network-on-chip (NoC), pose great challenges for thermal management. In order to keep system temperature within its thermal limits, the existing algorithms fail in effective load balancing, throughput improvement. This paper proposes an adaptive thermal-delay-throughput aware routing algorithm for vertically-partially-connected 3D NoC based on Stackelberg game routing algorithm (STAR). In the game, thermal, delay and throughput are mapped into leader and two followers, respectively. Firstly, the leader (temperature) set up the system, then each of the follower (delay and throughput) chooses an appropriate routing strategy in response to the leader. After a couple rounds of gaming, if the payoff of the leader stays unchanged when any single player changes its own strategy unilaterally, then the equilibrium is obtained, and the optimal routing strategy is found; therefore, the data packet can be transported across different layer to mitigate the thermal problem by considering the temperature, delay and throughput simultaneously. Experimental results on Noxim show that STAR reduces the standard deviation of routers temperature in a 3D NoC by 14.1–41.4% and the statistical traffic load distribution by 14.1–46.7% under various synthetic and traffic patterns. And the proposed STAR achieves 31.01%, 44.24% and 42.57% improvement in the global average delay, compared to three previous algorithms. And the STAR algorithm is comparable or better than the Reflect3D and LETHOR in terms of traffic load and throughput. The simulation results on different network sizes proves the scalability of the STAR algorithm.</p>