Well, I just tried a llama.cpp build (not execution of the program itself) and after a while half of the readings were above 80°C, sometimes even 90+°C (with ambient temperature of 26.5°C), e.g.:
I have also been running Jeff Geerling’s High Performance Linpack (HPL) benchmark with similar (or perhaps even slightly worse) results. In that case sometimes I was not getting crashes, but the final numerical consistency check (that is, the residual check) was failing, so the symptom of overheating was erroneous calculations.
FYI my board is installed in a Corsair 2000D AIRFLOW case without any case fans, of course, since I have to figure out how to wire them up, given that there aren’t any fan headers except the one for the main SoC cooler. As mentioned before, I also use a graphics card and a proper SFX power supply.
I’m using the Radxa AI Kit Case - and a small copper plate sandwiched between the CPU and heatsink (more context here - it seems like the AI Kit makes poor contact by default: Test of the AI PC kit ).
I’m thinking my issue probably isn’t heat. I did consider faulty memory as a possibility because it would usually happen while benchmarking large dense models (32B or 70B) but, although I haven’t run Memtest from a bootable, I did run a Memory Tester App within Debian (cant recall the name) for a while and that all seemed okay. Might be something to do with the CPU Scaling Governor or something - I think I recalled flicking that setting at one point and didn’t have it crash (commands are mentioned here: DeepSeek-R1-Distill-Qwen-1.5B on Orion O6 CPU ).
BTW, if you’re using the Orion for LLM’s, consider using ik_llama over llama.cpp. There’s some optimizations there that improve Prompt Processing around 2x on the Orion: Llama.cpp - Benchmarks
I agree that these temperatures don’t look right. They match what I initially got with the heatsink that comes with the AI PC kit that wasn’t touching the SoC well. So here I suspect a similar problem, maybe the heat sink is not properly mounted on the board or doesn’t touch flat. It would be worth disassembling it, inspect it, and reassemble it.
In my case , the fixations were too high for the heat sink to push enough on the SoC. Maybe it’s the same problem here. Or we could imagine a washer being placed on the top of the board, preventing it from pressing against the SoC. In the worst case, try with a 0.5mm thermal pad. If that fixes the problem it will clearly prove there’s an issue with the contact there.
Regardless of the AI PC kit, I think that standard heat sinks and fans are used on the board, including the one chosen in the PC kit, and that the SoC is simply not as thick as socketted CPUs used on PCs, and that might be why the one of the PC kit doesn’t touch well. But this doesn’t mean that the stock one is perfectly suited either. At least it worked fine for me, but it would be possible that it depends on tolerance or assembly.
The CPU in this SoC is clearly capable of doing 2.8GHz for some cores.
And IMO the CPU is the biggest bottleneck in this board, So any help it could get makes a big difference.
From my testing this board seems perfectly fine for general basic day to day work. (But that’s a very very low bar)
However this board essentially turns into a raspberry pi 5 duct taped to 2 raspberry pi 4’s if you try gaming on it or running software that utilizes few cores..
I personally do feel a difference going from 2.6 to 2.8ghz on 2 cores, So if this goes away i don’t see why anyone would want to “upgrade” their UEFI version.