I currently have an NVIDIA Quadro k620 GPU card (http://images.nvidia.com/content/pdf/quadro/data-sheets/75509_DS_NV_Quadro_K620_US_NV_HR.pdf) which I use in my research. I need energy consumption information for some OpenCL kernels that I’m running. Im using nvidia-smi but, according to the manpage, the reported power is for the whole board except when there are no INA sensors (which in this case it reports the power for the GPU chip only). I couldn’t find any information on which case this card fits.
My question therefore is: for this GPU, nvidia-smi reports whole board consumption or GPU chip only?
What’s an “INA sensor”? As far as I know, for all actively cooled GPUs (i.e. those with a fan), nvidia-smi reports the power consumption for the entire board including the fan. Which usually is what is relevant to users.
Please note that the sensors on these cards are not designed for high-precision measurements. You should assume an error of +/-5% unless NVIDIA’s specifications tell you otherwise.
[Later:] I checked various versions of the nvidia-smi documentation and could not find any mention of INA sensors.
Power Draw
The last measured power draw for the entire board, in watts. Only available if power management is supported. This reading is accurate to within +/- 5 watts. Requires Inforom PWR object version 3.0 or higher or Kepler device. Please note that for boards without INA sensors, this refers to the power draw for the GPU and not for the entire board.
You said that there is a 5% error, you meant 5 Watts (as described in the manpage)?
No, I meant 5%. I have never seen the specific language you are showing, with the +/-5W claim and the mention of INA sensors.
Honestly, I am very skeptical about that +/-5W claim, given that GPUs can draw anywhere from 30W to 300W (low end devices to high end devices). Electronic components usually have a percentual manufacturing tolerance. The website for the particular INA sensor you pointed to even has a diagram showing total error (%), which for most of the range indicates around 3% error.
Apparently, INA stands for instrumentation amplifier, an op-amp like circuit.
this manpage is from nvidia-smi version 384.81, from 2017.
Thank you for your help and clarification on the meaning of INA. If the power readout reported by nvidia-smi is for the whole board, I’m amazed with its energy efficiency (1W on idle, 9W running some opencl kernels), I guess I’m too biased from the times where GPUs were extremely power hungry.
I am CS guy and no expert on the matters, but best I understand over the past few years NVIDIA has not only optimized the power consumption of the GPU chips themselves (turning off power to unused portions in very fine-grained manner), but has also increased the efficiency of the on-board power-supplies (voltage regulators)on the cards and fine-tuned their power management software (automatic adjustment to clocks, including PCIe; voltage adjustments). This push for power-efficiency started with the Maxwell generation of parts.
You remember correctly that the older cards based on the Fermi and Kepler families were pretty bad power hogs, although in all fairness one should say that relative to the performance they provided they were still more efficient than standard CPUs by several times.