PT100

Cooling loops need at least 19 seconds to answer a 50-millisecond GPU power spike

DataCenterDynamics does the arithmetic on CDU response lag and gets 19 to 58 seconds against accelerator power steps of tens of milliseconds. That leaves the reliability question sitting unmonitored at the chip junction.