For completeness, here is some additional background information, so in case you think that it should be easily solvable you do not have to speculate or guess which obvious parameter we did not play with:
* we have redesigned the stencil three times to optimize solder paste amount * we have tried different soldering profiles * we have experimented with different solder pastes and even using none * the upper chip (RAM) is dipped into flux and lower (CPU) is placed into solder paste * we know that warpage of the DM3730 is the key problem because we placed a tiny snippet of Kapton tape under the middle of the DM3730 and then we did not see any shorts. But the RAM chip wasn't soldered any more at all. Because the DM3730 now did lift the RAM off from its top. But the DM3730 worked (without RAM of course). * soldering them to a stack before production is not possible because we have no holder for it * soldering the CPU first and the RAM in a second reflow step has been tried but is stress for the other components and sometimes has the same effect of shorts: testing the OMAP before installing the RAM, the OMAP asks on RS232 for pserial and we even can load a MLO. But after soldering the RAM on top we have had sometimes shorts under the OMAP. * we solder with a vapor phase machine which tries to heat up every BGA ball as uniform as possible and does everything with lower temperature (~235°C) than infrared/hot air (~260°C) * the main PCB shows no warpage * TI has OMAP4 application notes which describe the factors that influence warpage: http://www.ti.com/lit/an/swpa182c/swpa182c.pdf (page 56), unfortunately with clear measures. * the same document has an nice fishbone diagram what to think of and we have of course checked that. I hope this additional information helps other DM3730 projects not to going through the same hard learning as we had, starting with "others have build millions of omap devices, so it must be possible" and concluding that it must be easy or a single source of failure which can easily be fixed... But finally finding out that it is not that easy. BR, Nikolaus > Am 23.02.2017 um 08:09 schrieb H. Nikolaus Schaller <h...@goldelico.com>: > > >> Am 22.02.2017 um 15:49 schrieb H. Nikolaus Schaller <h...@goldelico.com>: >> >> >> But now the bad news. >> >> -1.: Only 12 out of 36 (*) GTA04A5 boards we >> have produced, are booting to Linux. The others either have a >> short circuit under the OMAP (which is the general PoP soldering >> problem we already know since the first generations of the >> GTA04) or the OMAP shows no response or does not recognise the >> µSD card for unknown reasons. > > Because we are a community to share experiences and some people > expressed (unfortunately not here) the question how Nokia was > obviously able to build the N900 or N9 and doubt the warpage > effect, let me write down some of the thoughts: > > * we do not know how much scrap others do or did produce (such info > is usually not shared with the public), so our 30% yield may > be equally good. Or may be worse. Since it is >0 cases you only > see the successful results in the wild. > > * most likely, they were allowed to scrap 1000 units at a total value > of let's say 300k€ to learn the process and then they were able to > build 1 million units at high yield. We simply don't have that > much money for experimentation with the production process. I.e. > we want to win champion's league but can't afford to do the complete > training. > > * there are subtle variations between the chips. Basically the OMAP > is a tiny 0.1mm PCB where the Silicon die is sitting in the middle. > The thickness of this PCB has a big influence because it is soft outside > of the Si-chip and the exact size of the Si-chip also has an influence > because it makes this PCB stiff in this area. TI did shrink the DM3730 > compared to the OMAP3530 and made the PCB thinner. So it is a completely > different thing wrt. thermo-mechanical effects. > > * the PoP-RAM chip sits on top and has a hard plastics case and does > not warp, but modifies the way the lower chip gets heat and how far > it can move until the upper side of the Si-die touches the bottom of > the PoP-RAM and starts to lift it off. Even the exact material of the > BGA balls of the PoP-RAM can have an influence how much they "pull upwards" > the DM3730 carrier PCB. Or press down the Si-Chip onto the PCB. There is another effect mentioned by http://www.ti.com/lit/an/swpa182c/swpa182c.pdf (page 56) the weight of the RAM chip can also have an influence how it presses and that could indeed be a factor between SAMSUNG and MICRON chips. > > * due to this special construction the BGA balls under the OMAP die > behave differently than the BGA balls where no die is above. The effect > is that we either have shorts or open joints in the inner or outer ring. > Finding a compromise between Scylla and Charybdis, where we have neither > shorts nor open joints, is the key to high yield. It looks (really looks > in X-ray) as we simply have not yet found the best compromise and now ran > out of money and components during this search. > > * it looks as if the iMX.6-PoP chips use a hard plastics case for the > processor - so that it can't warp at all and all BGA balls on the lower > and upper end have the same mechanical background. So one could say we > fight with a design weakness of the DM3730 package design. To be fair it > must be said that the PoP stack of the DM3730 appears to be thinner than > the iMX.6 which might be crucial to squeeze the PCB into a smartphone > case without making bulges. > > * and not to forget: we don't have 0% yield, like it would happen with > a simple software or hardware bug. Those are usually show-stoppers and > when fixed, it goes perfect afterwards. We have 30% yield. This means that > production is possible, but just too expensive for us to find and eliminate > the > final disturbing factors. So it is mainly a financial limitation we have. > If the community can afford to pay 1500€ for each GTA04A5, we can start over > tomorrow. So one could even develop a crude view that the bug is that the > community can't afford what we are doing. Well, I have no idea how to fix > such a bug... Would be nice to learn. > > * others failed as well with this or similar combinations. For example > the OpenPandora was initially planned to be built in the US by the same > production line as the BeagleBoard was built. But if you ask Pandora owners, > they will tell you about the pain when it initially failed. > > * we then suggested to ED to go to the production company we now have used as > well, > and the OpenPandora 600MHz (OMAP3530) was successfully built. When they > switched > to the 1GHz Pandora some problems came back again, but were solved again. > What we > also have to consider is that during this time people have changed and > machines > were upgraded which is a process that usually looses some knowledge or has to > be regained for a modified situation. > > * for the GTA04A3/4 we also had to fight the same issues at a different > production > site. And after a long learning curve it worked better. You may ask why we > didn't > choose them again. Well, they knew about the low yield problem and they would > have > asked for much more money for trying production again without warranty... > > So it is like soccer. You know all rules, you know the dimensions of the > goal, the > ball and the shoes, the playground and that the grass is green. You also know > your > players - but making a goal is not at all guaranteed by this knowledge. It > needs > training and luck. > > Hoping this big bunch of experience helps other projects to succeed, > Nikolaus > > _______________________________________________ > Gta04-owner mailing list > gta04-ow...@goldelico.com > http://lists.goldelico.com/mailman/listinfo.cgi/gta04-owner _______________________________________________ Community mailing list Community@tinkerphones.org http://lists.goldelico.com/mailman/listinfo.cgi/community http://www.tinkerphones.org