I did my custom carrier board for a Jetson Nano module with a DC-DC converter, providing 5V (and 5A) to the Nano module. My software uses a USB3 camera, powered by the same 5V. Power is clean and stable the whole time. The software processes every frame and the profiler shows that it uses around 50% of the time between frames.
But, if I set my voltage to exactly 5V, the software doesn’t work. It behaves as if the GPU was slower, it cannot process the data (although it can transfer it). If I increase the output to 5.2V, everything works fine.
I am talking about voltages measured directly at the pins of the connector for the module.
Looking at the output of tegrastats I can see the fields POM_5V_IN going from 1900 (sw not working) to ~2300 (sw working).
Does the speed of the Nano depend on the input voltage?
If you’re using the barrel connector, there’s some evidence that there’s a reverse-voltage protection diode there, and you need to provide enough to compensate for that voltage drop (5.2-5.35V)
This was recently mentioned in another thread.
So, I wouldn’t say that the “speed” depends on the voltage, but the “ability to run at the configured speed” may depend on the voltage, and if you provide 5.0V on the barrel connector, only 4.8V may get to the module, and thus it may perhaps not be able to run stable.
Thanks for your answer. I’m using my own board, not the development kit. Also, I’m measuring the voltages directly at the module itself. I’m using the module taken out of the development kit.
My software is quite demanding, I have to use it with jetson_clocks.
I just want to double check that this is not expected (and that the problem is in the power regulation of my board, even though I can’t see it).
It ended up being a problem with my DC-DC converter. It was too slow for the fast current peak when the USB and the application started, and for an instant the voltage went down to 4V (for a very, very short time). That was causing all the issues.
Yeah, that’s a real challenge with DC-DC converters. Either they turn into megahertz oscillators, or they become too slow to react to big power steps :-(
Careful feedforward/compensation RC network phase analysis allegedly can make it so that you don’t need to build one, measure it, and then build a second one that actually works. I’ve had some luck with that in the past, but it seems like a black (well, dark gray) art to me.