Skip to content

06/09/26 Broken AIO Cooler

Yesterday I finally got my dedicated local AI server into a state that I was happy with.

I had wiped the old machine, installed Ubuntu Server, got the NVIDIA drivers working with the RTX 3080, installed Ollama, confirmed models were actually running on the GPU, set up SillyTavern, made it accessible properly around the house, and got to the point where I could switch my main PC off and still use my own local AI from my phone.

It worked.

For one day.

I went to turn it back on this morning because I had the entire day planned around continuing the project. The next stage was supposed to be image generation, then video generation, transcription, voice generation and eventually a larger local knowledge system.

Instead I got:

No route to host

The server was not appearing on the network at all.

At first I assumed the issue was LUKS, because the system had been powered on for some time before I decided to attempt unlocking it, giving me that error.

The machine was not responding to ping. There was no ARP response. The router did not show it connected.

That meant this was happening before Ubuntu was even getting far enough to bring the network interface online.

The Monitor Comes Back Out

The entire point of this machine is that it is supposed to be headless, but I dragged a monitor and keyboard over to it anyway.

The first useful thing I saw was the ASUS POST screen:

CPU Over Temperature Error!
Press F1 to Run SETUP

Well.

That explained the networking problem.

The server had never reached Ubuntu. The motherboard firmware had stopped the boot because the CPU was overheating.

I powered it down and left it for about twenty minutes before trying again.

This time there was no immediate over-temperature warning, so I entered the BIOS and watched the CPU temperature.

It was already around 86°C.

Then 90°C.

This was while sitting in the BIOS doing essentially nothing.

Power off again.

The Cooler

The CPU is an Intel i7-12700K and the machine was originally built as a gaming PC.

The cooler is the original:

PCS FrostFlow 240 Series RGB liquid cooler

It is a 240 mm AIO: a pump/block mounted to the CPU, two coolant hoses, and a radiator with two fans mounted at the front of the case.

The obvious suspect was the liquid cooler.

The strange part was that everything appeared to be running.

The BIOS showed roughly:

CPU fan:    ~1820 RPM
AIO pump:   ~2200 RPM

So the pump was reporting a speed.

The radiator fans were spinning.

Nothing was visibly leaking.

The CPU block was firmly mounted.

The hoses were warm.

One hose was noticeably hotter than the other.

From the outside it did not look like a dead cooler.

Dust

The radiator, however, was filthy.

I clean the fans occasionally, but I realised I had never properly cleaned the radiator itself because it is awkwardly mounted behind the front fans.

For a while this looked like a promising explanation.

I went out and bought compressed air.

Then I spent a large portion of the afternoon dismantling the front of the machine and cleaning the radiator, fans and case properly. I put everything back together and booted into the BIOS.

Initially:

59°C

Then:

65°C
70°C
80°C

Still climbing.

So the dust was certainly not helping, but it was not the fundamental problem.

Thermal Paste

Next theory: perhaps the CPU block was making poor contact with the CPU.

I removed the block.

The old Arctic MX-4 thermal paste was still wet and fairly normal looking rather than dried into concrete, but at this point there was nothing to lose by cleaning everything and reseating it.

I cleaned the CPU heat spreader and the copper cold plate with isopropyl alcohol.

I still had an unopened tube of thermal paste from when the machine was originally built, so I reapplied thermal paste.

This time the initial temperature was significantly better:

49°C

For a moment that looked encouraging.

Then:

1 minute:  60°C
2 minutes: 65-70°C
            72°C
            73°C
            75°C
            77°C
4 minutes: 80°C

That was the end of the experiment.

A 12700K should not steadily march towards its thermal limit while sitting in the BIOS with a functioning 240 mm liquid cooler attached to it.

Cleaning the radiator and replacing the thermal paste improved the starting temperature, but the cooler still could not remove heat from the CPU at anything approaching a normal rate.

The Likely Failure

The AIO pump still reports approximately 2200 RPM, but RPM does not necessarily mean useful coolant flow.

Technically it is possible to cut open, flush and refill some AIO coolers.

I am not going to do that.

This is a four-year-old budget sealed liquid cooler sitting directly above an expensive CPU, motherboard and RTX 3080.

There comes a point where repair becomes stupid.

Back to Air Cooling

I ordered a:

Noctua NH-D15S

Noctua explicitly lists the motherboard as compatible, although the heatsink has to be installed in the correct rotated orientation to clear the ROG shroud.

The case supports a maximum CPU cooler height of 165 mm.

The D15S is 160 mm.

Close enough.

Once the liquid cooler is removed, the radiator goes with it. The two radiator fans can potentially remain as ordinary front intake fans.

So Much for My Sunday

This was not what I intended to spend today doing.

My plan had been to start turning the machine from "Ollama server" into something much more ambitious.

The eventual system I have in mind includes:

  • local LLMs;
  • a source/document knowledge system;
  • persistent project context;
  • local transcription;
  • local voice generation;
  • image generation;
  • video generation;
  • automated overnight content-production jobs;
  • a research/news monitoring system;

Instead I spent essentially the entire day:

  • network troubleshooting
  • firmware troubleshooting
  • thermal diagnosis
  • dismantling computer
  • cleaning radiator
  • cleaning CPU
  • replacing thermal paste
  • remounting cooler
  • watching temperature climb
  • admitting defeat
  • ordering a giant fan

There was a brief period yesterday where I had built an AI server.

Today I built a very clean broken computer.

Still, the day was not entirely wasted.

I now know:

  1. The remote-unlock/network configuration was not the problem.
  2. The motherboard's thermal protection works.
  3. The radiator desperately needed cleaning anyway.
  4. The CPU block has fresh thermal paste and good physical contact.
  5. The continued temperature rise after all of that gives me a very strong diagnosis of a failed/degraded AIO.
  6. The entire case is now clean before the new cooler goes in.
  7. I have learned considerably more about CPU cooling than I knew yesterday.

Most importantly, I discovered the cooling problem before putting the server under sustained image/video-generation workloads.

That could have been a much more expensive discovery.

The NH-D15S is due Friday.

Until then, the server remains unplugged.