Thermal throttling

Thermal throttling is a processor automatically lowering its clock speed when it reaches a temperature or power limit. The chip trades speed for safe operation, so a job keeps running but takes longer. NVIDIA GPUs pass through three thermal stages. Software lowers clocks above the maximum operating temperature, a hardware slowdown cuts them harder, and the GPU powers off at its shutdown temperature.

NVIDIA's RTX PRO 6000 Blackwell Server Edition lists a 1,852 MHz base clock and a 2,430 MHz boost clock. GPU Boost raises the clock as far as the maximum operating temperature allows, and the hardware slowdown cuts core clocks by 50%. Running nvidia-smi -q reports clock event reasons, including SW Power Cap, SW Thermal Slowdown, HW Thermal Slowdown and HW Power Brake. If every reason reads Not Active, clocks are running as high as possible. Separate counters record, in microseconds, how long each reason held clocks down. Intel processors cut frequency and power at Tjunction max, their junction temperature limit, usually 100 °C to 110 °C.

Judge a system by sustained clocks, not boost clocks. Run the real workload for its full length while logging clocks and the clock event reason counters. Treat any active thermal reason as a cooling fault to fix before deployment.

Sources

Last verified October 7, 2026.