Just a quick summary with some questions with what we know so far, maybe it helps finding ideas.
1. With the same rootfs and kernel version, 2GB RAM versions seem to run stable, 4GB RAM version don't. Besides the RAM, they use a different MLO/U-Boot and a different kernel (LPAE-enabled).
2. The maxed out value seem to make the unit freeze a lot more often.
3. A freeze is often accompanied with graphical glitches on the screen
4. But the serial log usually doesn't show anything. It simply freezes. Shouldn't a kernel panic occur when there are memory errors? Sure, if the error is THAT bad that it will instantly freeze the CPU, then nothing will happen - but how likely is it that Linux NEVER reports bad memory but instead freezes right away?
5. U-Boot has never crashed so far. I had memory errors with mtest on one PCB with the older sdram3-values, but I have only run the test there once. At least with the new values, all of them seem to run fine (tested for over 100 iterations)
6. According to the simulation using the IBIS files, the DRAM- and PALMAS-Connections are within the given specs of TI. So while a hardware error IS still possible, it should hopefully be not the reason. One thing to notice is that the improved PCBs (that are a lot better according to the simulation) don't seem to have improved at all compared to the older PCBs (which were out of specs according to the simulation). One should expect at least a bit of improvement.
As the 2GB RAM boards are working fine, we should check what's DIFFERENT with the 4GB RAM boards.
1. Obviously, the RAM Chips are very different (two chips in one, probably needing twice the bus speed, correct?) as well as their timings.
2. DTB-File used
3. Kernel (2GB RAM: non-LPAE). Do we even need LPAE? non-LPAE should support memory up to 4GB...?
4. Power usage - the double-die memory chips might use more power, maybe?
So, anyone has any ideas what to try? I'm ready to test everything ASAP :)