Finally got memtest to run, failing test 0, I think it's the first time I see that. It the same address that fail.

Well, I removed the fan and now I can run all the benchs in a row... Everything maxed out and 40MHz fsb... (did not manage that since yesterday after many try and different configuration test)
Yep, I just remembered that yesterday I began the benchs without a fan but I did find the cpu heatsink a little too warm to my taste and now that I think about it, it's when I stopped being able to bench...

Something in there like to be a little toasty! It's probably the limit of "stable" can be and somehow, heat (somewhere), help...

Rav wrote on 2026-05-31, 01:29:

Well, I removed the fan and now I can run all the benchs in a row... Everything maxed out and 40MHz fsb... (did not manage that since yesterday after many try and different configuration test)
Yep, I just remembered that yesterday I began the benchs without a fan but I did find the cpu heatsink a little too warm to my taste and now that I think about it, it's when I stopped being able to bench...

Something in there like to be a little toasty! It's probably the limit of "stable" can be and somehow, heat (somewhere), help...

With the RAM issues you're seeing and the fact that keeping the board cool has negative effects on stability, it's quite likely to be a loose pin on one of the UMC chips. I assume your board is one of the later revisions with the COAST module slot and the 'cache' being 2x rectangular chips by the memory slots?

On at least one of my M919 boards, the UM8881F chip by the memory slots had some loose legs and reflowing the solder on those pins to make sure they were all making good contact, resolved the no-post issue I was seeing unless the board was pressed in the right spot.

Your issue is less bad but it being dependant on a warm board suggests thermal expansion is making the loose contact bridge the gap when it's warm. If you have anything that's good for magnification, it might be worth going round with a needle/fine point tweezers to look for a loose leg. I'd check the UM8881F first since that's responsible for memory.

Thermalwrong wrote on 2026-05-31, 02:50:

With the RAM issues you're seeing and the fact that keeping the board cool has negative effects on stability, it's quite likely to be a loose pin on one of the UMC chips. I assume your board is one of the later revisions with the COAST module slot and the 'cache' being 2x rectangular chips by the memory slots?

On at least one of my M919 boards, the UM8881F chip by the memory slots had some loose legs and reflowing the solder on those pins to make sure they were all making good contact, resolved the no-post issue I was seeing unless the board was pressed in the right spot.

Your issue is less bad but it being dependant on a warm board suggests thermal expansion is making the loose contact bridge the gap when it's warm. If you have anything that's good for magnification, it might be worth going round with a needle/fine point tweezers to look for a loose leg. I'd check the UM8881F first since that's responsible for memory.

Yes, I was thinking about exactly that, thanks for confirming it to me. I have indeed the board you are describing (except mine don't have the fake cache chips, just solder pads to go nowhere... It's a 3.4F... I think)
I will make closer inspection tomorrow

I did tap gently on both the 8881F and 8886BF. It crashed while I was taking care of the 8886BF.
So it's probably that. Thanks!