Am 16.10.2017 um 23:09 schrieb Tony Lindgren <tony@atomide.com>:

* H. Nikolaus Schaller <hns@goldelico.com> [171016 14:03]:

Am 16.10.2017 um 22:10 schrieb H. Nikolaus Schaller <hns@goldelico.com>:

So I'll try to boot this board with 1400mV and compare memtester results...

And first result running at 1400mV:

memtester finds much less errors, but still finds a handful:

From what I recall I started hitting what seemed to be thermal issues
running memtester at 1.5V. To count out thermal issues, add a fan when
running the tests.

I have the impression from my tests that thermal influence is a small additional
effect to bring signals out of the expected borders. So it is not the real problem
but may make it works in some cases.

Here are new results:

A) VCDD-DDR3L is quite stable at the Palmas end. It fluctuates between 1.402mV and 1.396mV
except during some situations where DSI: omapdss DSI error:is reported (I have
no panel connected). Then it may go up to 1.412V. During that time(out) it may be that
there is no DDR activity at all as the OMAP may be running completely from cache in some
waiting loop.

Conclusions:
* Palmas regulation and voltage specifications are very precise (1-2mV!)
* the voltage drop happens between Palmas output and our DDR3L chips => power distribution network too weak => PCB layout issue

It was in the first PCB layout (V4.9) already years ago and just copied over and over,
because it was initially "good enough" and turned out to be a problem only after we
tried 4GB...

B) Next, I did measure the difference between Palmas output and one of the DDR3L-Chips (U3402).
Here, I see alternatingly 8 or 9 mV and 15-17mV.

This seems to be just "one rank" vs. "dual rank" power demand... Depending on what the memtester
is doing.

Conclusions:
* fully consistent with 2GB vs. 4GB problems

C) I did the same measurement for GND but there it is <0.3mV which is probably already
beyond precision of my equipment.

Conclusions:
* it is not the GND network but only VDD-DDR3L

D) after installing some additional VDD-DDR3L wires:


VDD-DDR3L is between 1392mV and 1403mV at U3402 - i.e. much more stable!

There is only one thing: it did not make all memtester failures go away.
One of the errors seem to indicate a problem on two specific data lines.

FAILURE: 0xa4c4cfc3 != 0xa4c4c7c3 at offset 0x0a4bfcc8.
xor diff:              0x00000800
FAILURE: 0x1e22c57e != 0x1e22c77e at offset 0x0c8f9070.
xor diff:              0x00000200
FAILURE: 0x6401c736 != 0x6401cf36 at offset 0x15460aa4.
xor diff:              0x00000800
FAILURE: 0xc64eef8e != 0xc64ee78e at offset 0x35342568.
xor diff:              0x00000800
FAILURE: 0x0440c7ed != 0x0440c5ed at offset 0x357e4cc4.
xor diff:              0x00000200

E) final test with 1350mV setup

did show many more errors - like before.

Conclusion: adding external wires doesn't help too much even if voltage seems to be more stable :(

So either this board is a little damaged/affected by rework (the other one works very well at 1.35V)
or power distribution must be optimized in PCB layout and my wires go to the wrong locations.

All this means we unfortunately still can not be 100% sure about the reasons of these remaining failures
with IM8G memory chips.

Any ideas for further cross-matching tests?

BR,
Nikolaus