Power and Thermals

Key Insight

Running multiple high-performance GPUs in a single workstation creates serious power and heat challenges that can silently degrade performance through thermal throttling — the chip automatically slowing itself to avoid overheating. Undervolting lets developers lower each GPU's voltage-frequency curve to reduce power draw (often by 50–100 watts per card) while maintaining stable clock speeds and nearly identical throughput. This project teaches how to measure real-time power consumption with nvidia-smi, monitor junction temperatures, and configure system-level power limits so that a multi-GPU setup runs continuously under heavy training loads without exceeding the cooling system's capacity or the power supply's rated output.