Will Zeus work accurately on a Jetson Nano?

[rileywhite-noblis]

Title.

[jaywonchung]

Hey @rileywhite-noblis, thanks for bringing this up.

We haven't thought about this but from searching, I'm afraid not yet. Zeus depends on NVML to read NVIDIA GPU power/energy stats, but NVML is apparently not supported on Tegra-based embedded GPU platforms. Instead, projects like Jetson stats seem to provide power information by reading from sysfs interfaces provided by Jetson's kernel.

I want to support Jetson platforms in the future. Can you briefly let us know your current use case? Are you running training or inference on the integrated GPU? Do you want to measure both CPU and GPU energy? How long would the unit of computation you would like to measure be?

[rileywhite-noblis]

We are trying to do an energy comparison between Jetson and neuromorphic accelerators. Currently, we are primarily interested in inference energy from the GPU. CPU energy consumption would be nice to have but not a necessity. In the future we do have interest in measure training energy consumption in comparison to neuromorphic accelerators, but it isn't a current need. We want to measure the energy consumption over as many samples in our dataset as possible to get the most accurate representation of the energy draw of the Jetson for our data/model. Because we are comparing to neuromorphic accelerators the data are time-series, so we need to know the per-timestep energy consumption averaged over all our data.

It would also be nice to measure the energy consumption of peripherals such as cameras to compare the energy consumption of the entire closed loop system when using a normal RBG camera with a Jetson vs a neuromorphic camera and accelerator. Not sure if this last bit is considered out of scope for Zeus though.

[jaywonchung]

Sounds good! Thanks. I've created a tracking issue here: #103. You can subscribe to the issue for updates.

We need to start by getting access to an actual Jetson device (including figuring out where to get the money 😅) so I will probably encourage you to look at the Jetson stats project for now, but we'll update you when support lands. I think extending support for the embedded land is super valuable.