Another test result with the second IMG 4GB board (which sometimes shows memtest errors):
I did measure VDD_DRAMat one of the decoupling caps of one DDR3L chip.
It fluctuates between 1.346V and 1.367V during the boot process. Heavily dependent on what the kernel is doing (e.g. DSS not yet active).
NOTE: the values shown here are low frequency voltages shown by my multimeter. They usually have a low-pass input filter. There may be high frequency drops which go much lower than 1.327V.
During console idle it was 1.348V.
Starting memtester did go down to 1.346V.
But for a short moment (stuck address test? I did not check time relationship) I did see it going down to 1.336V.
While running the random value test it was constant at 1.346/47V.
Then, voltage did go down to 1.330V for short moments and the random value test showed a lot of FAILURE:
The lowest value I could see on the multimeter was 1.327V during the Solid Bits test (which did not fail).
This and all other findings are consistent with a too high resistance between Palmas and VDD of the DDR3 chips. 1. the more active the memory is, the higher is the current demand, the higher is the voltage drop 2. this makes many boards fail as soon as DSS (around 7-9 seconds) is enabled (DSS adds constant DDR access through DMA) 3. higher voltage seems to "cure" it - at least partially 4. 1.5V is not significantly more stable - 1.5V has much higher current consumption and therefore more fluctuations 5. 4GB memory draws more current which makes the effect worse than with 2GB (or if we operate the 4GB chips in 2GB only mode)
So I'll try to boot this board with 1400mV and compare memtester results...
And I want to try to add some thick copper wires between Palmas output and DDR3L VDD. Fortunately these points are easily accessible at the decoupling caps.
If the board passes the memtest in either case (please cross fingers :) it proves the theory. And we finally have found out what the key reason is and can start to think about (cost effective) solutions.
BR, Nikolaus