Apologies if this is a simple question and my Google-fu has failed me but…
The docs for the Tesla K80 show that there is a “Zero-power idle” mode. In my mind that differs from being in P8 (idle) where each GPU draws in the region of 25W. 50W per card total is NOT zero, so this doesn’t seem like idle and zero-power idle are one and the same thing.
The only other place that gives explicit mention of this feature (and it may not even be done at the GPU end) is IBM’s LSF documentation - specifically there is a setting LSB_GPU_POWEROFF_DURATION. The docs say that it is possible to power off a GPU when it is not in use to save power (doesn’t say if this is a whole slot or a single GPU).
A cursory look through the NVML API shows some calls to read and set power states, spanning P0 to P15 (though the card seems to use P8 as the idle state), though no mention of powering off a GPU.
There is the “drain” mode available in Linux via nvidia-smi but that looks to prevent new things from being run on the card - the power meter shows no change when cards are put into “drain” mode if the card was already idle to start with.
Question is - how can one activate the zero-power idle feature of a Tesla card ?