Jetson compute utilities
The AI Accelerator compute box is an embedded NVIDIA Jetson Orin AGX 64GB (Tegra). Use the commands below on the compute module itself — keyboard and display connected, or SSH as user lab (password easybot). Default IP is 192.168.0.100 — see SDK.
For SDK work you usually run perception inside Docker, but CPU load, temperature, and power mode are easiest to read on the host before or while containers are running.
Quick reference
What you want |
Command |
|---|---|
Live CPU / GPU / RAM / temperature / power |
|
Interactive dashboard (if installed) |
|
Power mode (MAXN vs low power) |
|
Software / BSP version (2.x) |
|
Tegra L4T release line |
|
Disk space |
|
Memory summary |
|
Running containers |
|
Container CPU / RAM |
|
tegrastats — CPU, GPU, memory, temperature, power
NVIDIA’s built-in monitor. Updates once per second until you press Ctrl+C.
tegrastats
Sample line (values vary with load):
RAM 8234/64341MB (lfb 1234x4MB) SWAP 0/32170MB CPU [45%@2201,12%@729,...] GR3D_FREQ 99% cpu@48.2C gpu@46.3C soc0@46.8C tj@48.2C VDD_IN 45234mW
Field |
Meaning |
|---|---|
|
System memory in use |
|
Swap usage (should stay low during normal SDK use) |
|
Per-core utilization and clock (MHz) |
|
GPU utilization (high during FoundationPose / model compile) |
|
Temperature sensors in °C ( |
|
Approximate input power in milliwatts |
Log a short sample to a file (useful when reporting performance issues):
timeout 60 tegrastats --interval 1000 | tee ~/tegrastats_$(date +%F_%H%M).log
Inside the SDK container, tegrastats is also available (the host binary is bind-mounted by run_sdk_setup.sh):
docker exec -it isaac_pkgs_app tegrastats
jtop — interactive dashboard (optional)
jtop (jetson-stats) shows CPU, GPU, memory, power mode, fan speed, and temperatures in a single terminal UI.
If jtop is not installed:
sudo apt update
sudo apt install -y python3-pip
sudo pip3 install -U jetson-stats
sudo reboot
Then run:
jtop
The SDK container can read jtop’s socket when the jtop group exists on the host (run_sdk_setup.sh mounts /run/jtop.sock).
Power mode — nvpmodel
Jetson power modes cap CPU/GPU clocks and TDP. The SDK expects MAXN (maximum performance), especially when compiling SAM / FoundationPose engines.
nvpmodel -q # show current mode
sudo nvpmodel -m 0 # set MAXN
Typical modes on the AI Accelerator Jetson Orin AGX 64GB:
ID |
Name |
Use |
|---|---|---|
|
MAXN |
Normal SDK use, perception, model compile |
|
MODE_15W |
Low power — avoid for AI workloads |
|
MODE_30W |
Low power |
|
MODE_50W |
Low power |
If model restore fails after a compile, confirm MAXN — see Model restore failed.
Optional: lock clocks at maximum for a benchmark (reverts after reboot):
sudo jetson_clocks
Version and platform info
AI Accelerator software version (2.x) — preferred check from SDK:
curl -s http://localhost/universal-robots/platform/system/v1/versions | python3 -m json.tool
Underlying Tegra / L4T release (useful for NVIDIA support tickets):
head -1 /etc/nv_tegra_release
On 1.x boxes only, the REVISION field in that line helps pick an upgrade file — see Upgrade from 1.x.
General host utilities
These work like on any Linux system and are handy during setup and debugging.
CPU and processes
htop # if installed; otherwise: top
uptime # load average
Disk
df -h ~ # free space on home / data partition
du -sh ~/aia_sdk/ros/data/* # size of datasets, models, logs
Large recordings and checkpoints live under ~/aia_sdk/ros/data/ — see Clean up data files before deleting anything.
Docker
docker ps # running containers (isaac_pkgs_app, aia-core, camera, …)
docker stats --no-stream # one-shot CPU/RAM per container
docker logs --tail 50 isaac_pkgs_app # recent SDK container output
Network
ip addr show # compute box IP (default 192.168.0.100 on 2.x)
ssh lab@192.168.0.100 # from another PC; password easybot
ping -c 3 192.168.0.10 # robot reachability (adjust IP)
When temperatures look high
During FoundationPose, SAM, or first-time TensorRT compile, GPU utilization and tj temperature will rise — that is expected. If the box throttles or shuts down:
Confirm MAXN:
nvpmodel -qCheck airflow around the compute box (fan exhaust not blocked).
Capture a short log:
timeout 120 tegrastats --interval 1000 | tee ~/thermal.logCompare load with containers stopped:
docker stop isaac_pkgs_appthen runtegrastatsagain.