I’m using the Jetson Orin Nano Super to test the effect of running a typical LLM on it.
The ollama and llama3.2:3b was set up under the official tutorial.
The power adaptor is the original one, and I’m using MAXN SUPER power mode without any power supply warning. The bug is after interacting with the llama3.2:3b back-and-forth for several times, the output will suddenly become “GGGGGGGGGGG“ no matter what you typed the last time.
I made a 16GB swap on the SSD, however the phenomenon still persisted even though the RAM hadn’t been filled up.
In spite of all that has been said, the kit itself produced an unpleasant, consistent high-pitch sound (the Coil Whine) when running ollama. Is that normal?
*** Please note that this reply is generated by LLM automatically ***
Based on the provided context, I’ll address the issues you’ve encountered with your Jetson Orin Nano Super.
Issue 1: Output becoming “GGGGGGGGGGG” after interacting with llama323b
This issue seems to be related to the power management and thermal settings of your Jetson Orin Nano Super. Although you’re using the MAXN_SUPER power mode, which is designed to provide a balance between performance and power consumption, it’s possible that the system is still experiencing thermal throttling or power limitations.
To troubleshoot this issue, you can try the following:
Monitor system temperatures: Use tools like tmon or jetson_clocks to monitor the system temperatures and ensure they are within the recommended range.
Adjust power settings: Try adjusting the power settings using the nvpmodel tool to see if it improves the situation. You can try setting the power mode to MAXN or MAXP to see if it makes a difference.
Check for software issues: Ensure that the llama323b software is properly configured and optimized for the Jetson Orin Nano Super. You may want to try updating the software or checking for any known issues.
Issue 2: High-pitched sound when running ollama
The high-pitched sound you’re experiencing is likely due to the power supply or the system’s cooling mechanism. Here are some possible causes:
Power supply: The original power adaptor might not be sufficient for the Jetson Orin Nano Super, especially when running demanding workloads like ollama. Try using a more powerful power supply to see if it resolves the issue.
Cooling mechanism: The Jetson Orin Nano Super’s cooling mechanism might be causing the high-pitched sound. You can try adjusting the fan settings or ensuring that the system is properly ventilated.
Issue 3: Swap space and RAM usage
Although you’ve created a 16GB swap on the SSD, it’s possible that the system is still experiencing memory constraints. You can try:
Monitoring RAM usage: Use tools like htop or top to monitor the RAM usage and ensure that it’s not being filled up.
Optimizing software: Ensure that the llama323b software is optimized for the Jetson Orin Nano Super’s hardware and is not consuming excessive memory.
In summary, to fix the issues you’re experiencing, you can try adjusting the power settings, monitoring system temperatures, checking for software issues, and ensuring that the power supply and cooling mechanism are sufficient. Additionally, optimizing the software and monitoring RAM usage can help resolve the issues.
*** Please give the thumbs up if you get answers you like. Or provide feedback to help us improve the answer. ***
Yes, this bug appears each time I interact with llama3.2:3b using the ollama structure.
Moreover, when I open up the Open WebUI to interact, the “GGGGGGG“ hell will immediately appear when I interact with it for the second time (back-and-forth round 2). Next, if I start a new session, the “G“ hell will appear earlier (half way the first round! ) .
This unbearable phenomenon and the coil whine have been going on for three days (exactly till now since I first received my jetson and started using the GPU on it to inference LLMs! )
At first, I wondered that this might be caused by the using up of the RAM, so I swaped a 16GB swap on my SSD, but the bugs persisted.
My power supply source is stable, and I was using a 256GB high-speed SD card as the system card.
I’m facing the exact same issue using JetPack 6.2 and running any model in ollama. I have also followed the official tutorial and do not user docker but ollama is directly installed.
This issue appears on a Jeston Orin Nano Super I have just recevied. (I’m using an SSD in my case)
The result could be “GGGGGGGGGGGGG” or “3333333333333” and also see “CONTROL_36CONTROL_36CONTROL_36CONTROL_36CONTROL_36”
Well I thought that’s a hard fault, so yesterday I returned it to the vendor.
Maybe they will later deliver me a new one to test.
But one more detail to ask, which is the high-pitch coil whine when the GPU is used. Have you ever noticed that? If this phenomenon also occurred on your jetson, the terrible sound might be normal.
I have finaly be able to use ollama without any issue on the Jeston using Docker or Jeston containes. Looks the issue is specific to ollama directy installed on the Jestion.
Yes, the fan makes some noise when there is a high load but nothing really wrong for me.
In my case the output is 3333333 of a repeated < unused31>.
Good to hear that the docker method seems to solve issues. Is there a clear step by step instruction for this?
I do hope people are working on fixing the issue, because running ollama natively should not be any big task as weaker hardware copes just fine (al be it with slower performance)
I have followed the instructions on https://www.youtube.com/watch?v=R0PjKr4d-gU to get a docker running ollama, but I still get the < unused31 > repeated output for a longer prompt (that works on weaker hardware)
The other problem is that when the llms do run they make weird statements that hint at them ‘forgetting’ parts of what was input or not applying things they ‘know’. And again, compared to the same models running on another system, these problem occur on only on the jetson, and not in similar circumstances on the other (weaker) hardware like a Pi5.
And finally, installing a slightly bigger model that is stated to work on the hardware it gives errors both natively and in a docker
update, I decided to try a clean install of only the official ubuntu image and only the steps from the working instructions.
Unfortunately, it still gives the same error although the llama3.1 I also tried just started to output rather random words and pars of words it tried to organise into sentences
Is it solved? Same issue happend on my Orin. I tested qwen2.5:7b qwen3:8b llama3:8b deepseek-r1-1.5B, they all have this issue with a long prompt. The only model works is gemma3:1b for some unknown reason
No, this has NOT been solved (and keeps happening with multiple models, both large and small). Thanks for keeping the thread alive.
I have not seen a solution yet and multiple models from small ones to big ones eventually stop working. On my current install I have not set any swap file, so caching issues there are not at fault.
Docker does not solve it and uses extra memory, which might even complicate things.
