Thermal Aware Intelligent Scheduling Framework for High Density Artificial Intelligence VLSI Chips

The rapid advancement of artificial intelligence (AI) workloads has led to increasingly dense VLSI chip designs, demanding intelligent thermal management to prevent performance throttling and hardware failure. Traditional scheduling models often fail to accommodate real-time thermal dynamics, resulting in elevated peak temperatures and suboptimal energy efficiency. This paper introduces TAIS-Net (Thermal Aware Intelligent Scheduler Network) which is a new framework combining a spatio-temporal workload profiling approach, predicting rare-event thermal spikes, and diffusion modeling with neural ODEs to optimize task placement in real-time. The suggested system has a hybrid neuro-symbolic reinforcement learning agent to maximize computational throughput but with safe thermal margins. Experimental findings on ten high-density AI benchmarks show that TAIS-Net decreases average core temperature, peak temperature and thermal spike events by up to 6.2C, 5.1C, and 68.3 % respectively when compared against traditional schedulers. Also, TAIS-Net is even more efficient in terms of energy (12.4% higher) and inference (6.5% higher) than deep RL-based scheduling baselines. The framework is characterized by a total success in the scheduling with 94.3 %, which is determined by the successful placement of thermal-safe tasks and minimum throttling. These findings confirm that TAIS-Net is a promising future trend in AI chip runtime controllers. The model is cross-workload and cross-chip generalized, which provides scalable and dynamically adaptive thermal-aware scheduling.