This page documents the different ways we can measure power consumption for the Jetson Nano 2GB.
Questions we want to answer:
nvidia-smi supported on the Jetson Nano? This approach follow the NeuralPower paper.No. It is not possible to use nvidia-smi on a Jetson according to this as it requires a PCI bus. Jetsons have an integrated GPU (iGPU) which is wired directly to the memory controller. Only a discrete GPU (dGPU) on a PCI slot can use that utility.
tegrastats tool.We also cannot use tools such tegrastats because the 2GB version has no INA3221 power monitor.
https://forums.developer.nvidia.com/t/tegrastats-doesnt-display-power-usage-for-the-jetson-nano-2g/161410
Further confirmation from Nvidia forum moderator here. The 2GB version is not supported to do power measurement.
This is most likely the only option we have with the 2GB version of the Nano. This paper uses LynSyn Lite Board which is a power-profiling tool that was designed for monitoring the power consumption of single-board compute such as the Jetson Nano.
The LynSyn Lite Board seems to be a fairly sophisticated board. Suggestion is that we can probably use a usb-c power meter to measure the power.
Ideally we will want a power meter that has logging capabilities that tells us power used at certain time so that we can generate something like this.
We also don’t have to worry about the delay cause by the power logger as suggested in the Profiling Energy Consumption of Deep Neural Networks on NVIDIA Jetson Nano paper as we are not starting a separate process for logging power.
However, as the paper suggest, the inference time that the sum of layer-time from the profiler is unequal to the total inference time. This is something we will need to figure out how much delay is there when we start the experiment.