Jetson power: how much does the power mode change the time per sample?¶

A Jetson module runs at different power levels. The same code, on the same board, takes a different time per sample in each mode. This page states the modes of each Orin module, the commands to read and set the mode, and how to measure edge.edge_bench --smoke in each one. The measurements table is empty: it is filled with the numbers of a board, not from a desktop.

1. The modes of each module¶

The table is the one of the NVIDIA developer guide, Platform Power and Performance. The number in parentheses is the id used by nvpmodel -m.

Module	Modes (default marked)
Orin Nano 4 GB	10W (default, 0), 7W_AI (1), 7W_CPU (2)
Orin Nano 8 GB	15W (default, 0), 7W (1), MaxN (2)
Orin NX 8 GB	MaxN (0), 10W (1), 15W (default, 2), 20W (3)
Orin NX 16 GB	MaxN (0), 10W (1), 15W (default, 2), 25W (3)
AGX Orin 32 GB	MaxN (0), 15W (1), 30W (default, 2), 40W (3)
AGX Orin 64 GB	MaxN (0), 15W (1), 30W or 35W (default, 2), 50W or 60W (3)

The memory and the TOPS of the modules are on the Jetson modules page.

2. Read and set the mode¶

sudo /usr/sbin/nvpmodel -q
sudo /usr/sbin/nvpmodel -m 0

The first command prints the current mode; the second sets the mode with the id in the table. The ids are per module: id 0 is the default on Orin Nano 4 GB and 8 GB, MaxN on Orin NX and AGX Orin. sudo /usr/sbin/nvpmodel -q --verbose prints the mode and the clocks.

3. Watch the board while the benchmark runs¶

sudo tegrastats --interval 1000

tegrastats prints one line per second with the power, the temperatures and the clocks of the module. Run it in a second terminal while the benchmark runs: the time per sample and the power of the mode are read together.

4. Measure the benchmark in each mode¶

python -m edge.edge_bench --smoke

edge.edge_bench --smoke runs every pipeline on a small synthetic stream and prints, for each one, the median, the interquartile range and the p99 of the time per sample. Run it once per mode:

for m in 0 1 2; do sudo /usr/sbin/nvpmodel -m $m; python -m edge.edge_bench --smoke; done

The loop sets each mode and runs the smoke benchmark. The id must be one the module has: on Orin Nano 4 GB the ids are 0, 1, 2; on AGX Orin 64 GB they are 0, 1, 2, 3.

5. The measurements¶

Module	Mode	Pipeline	Median ms/sample	IQR ms/sample	p99 ms/sample
to be filled with the numbers of a board

The table is empty. A number written here without a board is a number invented, and this page does not invent numbers. The script that fills it is edge/board_selftest.sh, which runs the smoke benchmark, reads the model from /proc/device-tree/model, and writes the JSON with the model in its file name.

Details¶

References¶

NVIDIA, Jetson Orin Nano, Orin NX and AGX Orin Series: Platform Power and Performance. URL in section 1. The table of the modes is copied from it.
NVIDIA, Jetson Modules. URL in section 1. The memory and the TOPS of each module.
edge/edge_bench.py in this repository: the benchmark the page runs.

Choices stated explicitly¶

The page states the commands and the table, and leaves the measurements empty on purpose: the numbers come from a board, and no board is available here.
The loop in section 4 sets the modes in order and runs the smoke benchmark for each. It does not restore the previous mode at the end: the last mode stays. A script that restores it is edge/board_selftest.sh for the single run, not for the loop.
The mode ids are per module, and the table is the source. A module whose id is not in the table is not covered.

Development diary¶

The first version of the page had a table with three rows filled from a desktop run of edge.edge_bench --smoke. The numbers were the desktop's, not a Jetson's, and the page does not compare a desktop to a board. The table is now empty, and the script that fills it is edge/board_selftest.sh.