Counter-intuitive solution to problem with Taiyo BT/WiFi module - but why?

[VK3DRB]

I recently did a lot of redesign of an 8-layer PCBA that someone else started, adding a few features, removing others, and making many improvements on what was already there. We have about 30 samples of a new PCBA. In about 10 of the PCBAs, the WiFi (Taiyo-Yuden WYSBHVGXG) drops out intermittently, on some boards more often than others. Three of those PCBAs drop out within a minute or two of initail connection to a WiFi router. The host MPU, an ATSAMA5D27 (Linux driver for Marvel 8887), issues a reset to the module via the nPowerDown signal, because it was not getting a response from the SDIO which has its logic powered by the module's internal 1.8V buck regulator. I did not change the legacy design on this WiFi except for one blunder by the previous engineer (he was transmitting the RF straight into the ground plane, rather than the antenna!), but I did notice he did some other weird stuff, like tying the JTAG-TMS line to the 1.8V rail. But he did copy the reference design on the output of the internal 1.8V buck regulator inside the module (filtering the 1.8V output with 4 x 10uF caps + 2 x 100nF caps).

Taiyo suggested I remove the JTAG-TMS line and just leave it floating and to add more capacitance (making it 6 x 10uF caps) to make this reliable. (I somewhat disagree in that leaving one 100nF in there is good for high frequencies.) Cutting the JTAG-TMS line from 1.8V is impossible on this BGA design. But in any case I changed to 6 x 10uF caps and the WiFi ceased working altogether. I thought backing off the capacitance until it starts working. This is counter-intuitive. I noticed the less capacitance, the more reliable this WiFI becomes. Backing off the capacitance to only 2 x 10uF + 2 x 100nF, the three modules stayed connected to the wireless router the entire weekend without a dropout! Now the 1.8V was always OK at around 1.78V, with about 140mV switching ripple. The ripple level does not change with capacitance. So the 1.8V rail will be a low impedance output, and the extra capacitance could be pulling in extra current from the internal buck converter during filtering the AC. Other than that I have no idea why less capacitance makes this thing completely reliable. I suspect all 30 boards will have a borderline problem that can be worked around by reducing the capacitance. Maybe the buck regulator switches into some mode during startup where too much current drain at startup makes it unstable, when the TMS line is pulled high. I have done a re-spin on the PCBA that has the JTAG not connected and a few other improvements. The other voltage rail 3.3V is clean. I just wish the previous engineer did some basic testing early in the peace, but he obviously did not.

Does anyone have any idea why this device would be so reliable with less capacitance? Why he higher capacitance would make this module unstable, in contradiction to the Japanese vendor's recommendations? Has anyone used this module before with such issues?

In a new variant of the PCBA, I do not use the Taiyo device primarily because of cost of the device and extra cost of regulatory testing for FCC and CE RED. I am instead using the brilliant little Sierra Wireless BX3105. Tiny. low cost and works a treat. But it is too late to change the Taiyo module on the other variant because of project schedule impact.