I'm performing an analysis of power consumption of CNN ( with the trtexec tool ) on Jetson TX2, varying the Host and Device frequencies.
In the past I have performed this power analysis through a multimeter, in which I measured the current that flows from my lab power supply ( set at 19 Volts ) to the carrier board. Then through the Ohm law, I algebraically obtained the power consumption of the carrier board. Finally I have subtracted from this consumption, the power consumed when the SoC is in idle. For sure this is not an optimal methodology but I guess that it give an idea of the power consumption of the TX2 SoC.
Recently I'm trying to do the same, not with the multimeter, but through the INA3221 sensor inside the board. I created a script that read the input power of the SoC rail: in_power1_input.
The problem is that this two results are very different: about 4 Watt with the multimeter and about 0.9 Watt with the sensor. I am not an expert in electronics and power consumption of embedded systems, but I think that 0.9 Watt at the maximum GPU and CPU frequency is too low.
Maybe the SoC rail does not consider the GPU consumption? or the INA sensor does not capture something that in reality consumes power or something similar?