I was away for a holiday, so I was catching up on this thread, seeing very interesting comments and results.
I have replaced the LM with some thermal pas I had laying around, and I have 16k+ results again in CB r23, results I had only seen the initial runs just when I got the replacement board.
I made some pictures of when I removed the LM.
I used a bit of brainpower, I ran the machine hot with CB R23 and flipped it and let it cool down so I could easily remove it, and that worked. Could be a useful tip for anyone looking to remove the heat-sink and do some servicing on it.
Most of the LM ran to the sides and a little layer was left touching around 50 % of the die.
You can see the peel I got in the pictures.
I think Framework is not using enough lm to account for the out flow to the sides once it liquefies.
This would explain initial results doing great, but once the LM runs thin and “leaks” to the sides you get worse thermals and or performance.
So I ran new benchmarks after I replaced the LM with the thermal paste. I noticed 2 things, my max temp on the sensor4/core4 didn’t reach 100c, instead it touched 97c. The big delta is still there. The fan ramp was much slower, and ran longer at lower rpm (from what I could hear).
This leads me to believe the plate that is attached to the head-sink is making sub optimal contact at that exact space where the core is.
I read LM could damage copper easily so I understand there needs to be another metal in direct contact but something is not right here.
I have ordered some new thermal paste, liquid metal and pads from thermal grizzly to do more testing. I also need more paste to redo the dGPU so I needed more thermal paste anyway, just added the rest to the order, I can always use it on other machines I have.
I might also decide to lap the heat-sink but I want to do more testing first. The thermal paste I have now is at least working and makes the fan run at lower rpm and or less frequent.