Thermal Management and Energy Efficiency Optimisation for High-Performance GPUs

Abstract

The rapid growth of artificial intelligence computing has substantially increased the thermal load of modern graphics processing units (GPUs), making thermal management a critical issue in high-performance computing systems. This paper summarises GPU thermal-management technologies across multiple scales, ranging from chip-level cooling components to data-centre thermal systems. The discussion covers the mechanisms of GPU heat generation and their relationship with power consumption and thermal performance, followed by a comparison of major cooling approaches, including air cooling, direct-to-chip liquid cooling, and immersion cooling, in terms of cooling capability, energy efficiency, operating stability, and practical applicability. Recent advances in thermal interface materials, microchannel heat-transfer structures, and intelligent thermal-control strategies are also reviewed. However, increasing power density and system integration continue to present challenges in balancing heat-removal capability, energy consumption, reliability, and implementation complexity. Overall, the transition towards higher-power and higher-density computing requires coordinated optimisation of cooling hardware, system architecture, and intelligent thermal management to improve the efficiency and reliability of future Artificial Intelligence (AI) computing facilities.