Thermal management is a critical challenge in modern computing systems, particularly in high-density architectures such as 3D integrated circuits. Computational workloads generate localized hotspots that degrade reliability, accelerate semiconductor aging, and increase the risk of physical damage. Conventional sensor-based thermal monitoring suffers from latency, limited spatial resolution, and inability to detect transient thermal events, often leading to delayed or ineffective mitigation. Additionally, temperature behavior is influenced by material properties, workload characteristics, and cooling efficiency, resulting in a non-linear relationship with energy consumption. This often causes suboptimal thermal control, unnecessary throttling, and increased cooling costs. To address these limitations, this paper proposes a fully software-driven, predictive temperature modeling technique using hardware and software performance counters. Machine learning models such as XGBoost are employed to proactively estimate temperature trends without relying on physical sensors. Achieving 86–99% temperature prediction accuracy across diverse workloads, the model enables dynamic, fine-grained thermal control, improving energy efficiency and mitigating hotspots in dense computing environments.