The Ware for March 2025 is part of a Wekome WP-U157 “67 watt” GAN power supply.

This particular unit had overheated and let out the magic smoke, so I decided to take it apart to understand what was going on. There were a number of engineering issues with the design; the most prominent of which is that the gap filler pads used to improve thermal conductivity had settled, and were no longer making intimate contact with the case. The small air gap meant the electronics were effectively cooking in a thermos.

I think Kienan basically nailed it (congrats & email me for your prize) – as with most Chinese products, the make/model number is just a throw-away layer of “business paint” and I suspect the actual OEM for a whole series of similar modules is Shenzhen Goodwin. As indicated by Kienan, the safety margins of this product are inadequate. I had picked up a couple of ~\$90 GaN chargers previously and was quite happy with them; this charger was available in Hua Qiang Bei for about \$30, so I decided to give it a try. It works, for small values of “work” — if you run it at anywhere near the rated capacity for more than a half hour, it’ll eventually shut itself down, presumably due to overheating.

I could believe it has no explicit thermal protection, and instead, perhaps they were relying upon the built-in thermal shut-offs of the regulator ICs to prevent damage. Why waste money on two safety mechanisms when you already have one? This is reminiscent of a “hint” I once saw about implementing battery cutoffs for shipping and long-term inventory storage…instead of adding an explicit FET to cut off a lithium battery to reduce leakage current during inventory storage, “here’s one weird trick” you can use … force the battery into overcurrent protection mode by shorting it before shipping the product. This effectively causes the battery to disconnect, minimizing leakage during transportation; the cutoff resets when you plug the device in and try to charge it, thus allowing you to save the cost of the cut-off transistor. Another motif I’ve seen is “just charge your batteries with a bad 5 volt supply and let the overvoltage protection on the battery act as your regulator”. Clever ideas, in a way, but still: shenanigans.

The charger itself contains three circuit boards. One is a high voltage backplane, that also serves as a heat sink. Line voltage enters this board and is routed through a bridge rectifier plus some filter/PFC caps. Then, there is a module that has the actual GaN device (Meraki MK2789CDG) which goes from the rectified AC line voltage down to 20V DC @ 3.25A, achieving an efficiency of about 93%. At 65 watts, that means about 4.9 watts of heat is generated by this stage alone (not counting the upstream bridge rectifier and subsequent down-regulators), which is already a lot of heat for a small volume without forced air cooling – more than a Raspberry Pi 4. Most of the internal volume of the power supply is filled with the isolation transformer integral to this circuit.

After the AC-to-20V DC stage, there is a set of conventional silicon DC-DC buck converters, which is the board shown as the ware for last month and repeated above for your convenience. The board is parallel to the “face plate” of the power adapter, so space is at a premium. It takes in the 20V from the tabs shown at the top of the image, senses the USB-C PD mode and multiplexes/regulates power to the three ports (2x USB-C and 1x USB-A). This adds additional losses to the supply, but in USB-C 20V PD mode, I would imagine the sync-buck FETs should be operating in a pass-through mode, so there would be no switching losses and you just have the resistance of the (in this case always-on) sync-buck FET + in-line inductors.

Oddly enough, most of the heat seemed to be coming from the face plate of the supply, and not the sides, so something wasn’t working right. I didn’t try to test it after cracking the whole thing apart, but my best theory is that maybe because the gap pads had failed, heat from the GaN core was only really able to escape through this regulator board. The resistance of the transistor and inductor both have positive temperature coefficients, which maybe combined with a lack of an explicit thermal shutdown could lead to a thermal run-away situation where the devices eventually burned themselves out. Datasheets on the YC-series FETs would have been enlightening, but I strongly suspect they don’t achieve the 2 milli-ohm rating hinted at by the sole hit returned by an internet search.

The Ware for March 2025 is shown below.

I was just taking this thing apart to see what went wrong, and thought it had some merit as a name that ware. But perhaps more interestingly, I was also experimenting with my cross-polarized imaging setup.

This is a technique a friend of mine told me about where you put a polarizer on the light source, and then put another polarizer 90 degrees from the light source in front of your imaging lens. As you can see from the comparison above, the contrast on the part numbers is greatly enhanced — at the expense of losing most of your light. What isn’t obvious from the two images is the one on the left required a much longer exposure because 90+% of the light is not reaching the camera due to the cross-polarization.

My understanding of the theory is that with this technique, light that hits textured surfaces is more scattered (and likewise the polarization is more randomized) by the rougher, laser-marked regions of a chip label, while light bouncing off of smoother surfaces is attenuated because the polarization is more preserved (and thus blocked by the cross-polarized filter on the lens). So it’s less of an enhancement of the laser marked regions and more of an attenuation of everything that’s not.

The solution to the attenuation problem in our modern era is to just throw more light at it and mount your camera on a stand so you can afford a longer exposure time. Fortunately white LEDs are really cheap and ridiculously powerful, and I have some extra COB-based ring light emitters laying around the lab. The technique doesn’t scale to large boards as easily, but it just so happens this particular ware fits nicely in the field of view of the illuminator + camera.

The whole setup is pretty crude…just a cheap polarizer film that was rough-cut with scissors, and taped onto a $10 COB ring light I picked up some time ago in the Shenzhen markets that screws directly into the accessory threads of the objective on my benchtop microscope. Works OK, tho!

The Ware from last month is the main board from a Lego Duplo Steam Train. As predicted, this was a much easier one to guess. Congrats to MJS for naming it (with a margin of just half an hour ahead of Will), email me for your prize! And again, thanks to spida for contributing yet another ware.

Here’s the Ware for February 2025:

Thanks again to spida for contributing yet another guest ware! Hopefully this one is a smidge easier to guess compared to last month’s.

The ware for January 2025 is the Gavilan SC laptop motherboard. The Gavilan laptop is one of the first portable computer designs, announced in 1983, at a 2024-equivalent price of $12,400. However, the company only survived for one year, per Wikipedia:

Owing to a rigorous overhaul of the design of the laptop, the company missed its initial shipment deadline of December 1983, with the first several dozen units shipping instead in April 1984. Early units were fraught with technical issues, prompting more tweaks. Mass production and sales did not commence until June 1984. By this point, a major distributor of the Gavilan computer had filed for bankruptcy and was forced to pulled out of their deal with Gavilan. In late 1984, Gavilan Computer Corporation themselves declared Chapter 11 bankruptcy with cash flow problems. The company ceased operations in 1985, having only shipped a few thousand units of the Gavilan SC.

The number of blue wires on the main board is pretty consistent with “fraught with technical issues, prompting more tweaks”, and the history of the company reads like a botched Kickstarter from 2010, except it happened in 1984.

I was a little surprised at how difficult it was to guess the nature of the ware. I think contestants were reasonably thrown off by the inclusion of the modem. That would be a pretty forward-thinking feature to be designed into a laptop mainboard back in 1984. It’s also puzzling why one would preferentially integrate the modem over say, the RAM/ROM on the mainboard: generally there is some bare minimum memory required for a system to even function, whereas a modem might benefit from modularity for e.g. SKU diversity and/or future-proofing against improvements in modem hardware.

That being said, the company didn’t survive. Perhaps it is due in part to poor integration decisions from the get-go, yet contestants were viewing the ware through the lens of good decision making. Personally, I have found some of the hardest wares to guess to be poorly designed wares, as the final form is a bad fit for the function.

After reading through the comment thread a couple of times, I find it too hard to assign any one person a clear winner. A shout-out to FETguy for the explanation of the crinkly solder mask: I think the SMOBC vs HASL answer is correct – turns out I probably knew the answer to this once upon a time, but I had forgotten it with old age. Unfortunately, there wasn’t a tie to break in the competition – I think none of the guesses got close enough for me to declare a winner, despite all the thoughtful commentary. Thanks to everyone who participated!