Thermal Design Power: What the Datasheet Doesn’t Tell You

TDP is the single most quoted number on any chip’s datasheet and one of the most commonly misread. It is not a promise that the part will never draw more power than that figure — it’s the number the thermal solution (heatsink, fan, vapor chamber, whatever’s attached) is designed around, under a defined ‘typical’ workload the vendor chose.

Modern parts, including several in our own lineup, deliberately exceed their nameplate TDP for short bursts, because most real workloads are not sustained at 100% utilisation — a burst above the rated figure for a few seconds, while a heatsink’s thermal mass hasn’t caught up yet, costs nothing in reliability and buys a customer better perceived responsiveness. Onyx does exactly this: its 6.5W typical figure is a sustained number, but it will happily pull closer to 9W for the first several seconds of a heavy inference burst before its power management throttles back to the sustained figure.

What this means practically for anyone designing a board around one of our parts: size your thermal solution to the sustained figure for continuous operation, but make sure your power delivery — the actual voltage regulator and decoupling — is rated for the transient peak, not the sustained average. We’ve seen more field failures from undersized power delivery reacting badly to a legitimate transient than from any actual thermal runaway. The datasheet’s absolute maximum rating exists precisely to tell you where that transient ceiling is; treat it as a power-delivery spec, not just a thermal one.