I’m running quantized detection models on my Jetson Orin Nano Super Dev Kit with JetPack 6.2.
Current Status:
Workload: 2x CSI cameras working at 8FPS, 640x640 resolution.
Power: Active consumption is 6.2W, but Idle consumption is 4.7W.
Inference only adds ~1.5W (which is great), but I don’t understand why the board draws so much power just sitting idle.
What I have tried:
Using the default 7W Power Mode.
Creating custom power modes (reducing clocks/cores), but this hasn't helped much.
My Goal: I want to reduce this 4.7W idle baseline. My application is minimal and strictly needs only:
1x Ethernet port
1x USB port
2x CSI camera ports
Questions:
Is there a recommended way to reduce this idle power?
Should I modify the Device Tree (DTS) to completely disable unused modules (WiFi, Display, Audio, extra USBs)? If so, which nodes are safe to disable?
Are there specific hardware modifications (e.g., disconnecting specific peripherals) recommended for this carrier board?
I managed to reduce consumption in idle mode to 2.5W, changing SSD to microSD, I turned off the display, audio, pcie in Debice Tree. I increased the temperature at which the heat sink turns on. In the idle mode, the consumption is already satisfactory, from 4.7W managed to reduce to 2.5W. However, in detection mode with a model with a size of ~20 million parameters in FP16, with a resolution of 640x640 and 8FPS using DeepStream, consumption decreased by about 0.5W (currently around 6W) compared to what was before optimisations, which is much less than in Idle mode. Could the reason be the high consumption of microSD in operational mode? It does not save anything on the card except system logs. Is there any option to reduce energy consumption under light load, I tried my own energy modes, reducing the number of cores and frequencies but it didn’t help too much.