MicroSD card FAQ

Flash is frightening.

I wrote a transactional flash file system (well, object store) in the early 90s. It used NOR flash, which allows you slam individual bits from 1 to zero in any order. Basic strategy for doing a write: Write the payload, verify it, then clear some "commit" bits, and have enough infrastructure bits around that you can handle commit failure from stuck-at-1 errors. My failure model was probably pretty stupid, but it seemed to work well and I never heard of someone losing data. Of course, flash parts were a lot clunkier back then.

Fast-forward 20 years and they're playing terrifying games with those flash cells. I worry that the people writing the firmware are willing to trade off reliability for capacity or speed. Or they make crazy risks ("Hey, this window is only for ten milliseconds, what are the chances?"). I know too many lame firmware engineers to be really comfortable.

The failure modes are not block-by-block now, either. The whole device can go dead just because the firmware lost a lookup table to a bad bit. Something like this very clearly happened to Intel a couple years ago, and I know of other manufacturers who weren't very careful with their firmware quality (along the lines of "If we bump the version number on every fix, this will just make our customers nervous." Wow).

I'd love to see an eMMC spec that:

- Allows raw access to the underlying flash (useful for data recovery, in any event)

- Provides metadata describing the treatment of that flash (how many reads / erases, leveling requirements, geometry info, etc.)

- Provdes access to ECC engines, buffers and other support hardware (to reduce bus traffic and host computation)

... then we could push the data management into a FOSS layer in the host, where people can inspect it, fix it, and improve it.

The combination of a hot commodity, pressure to make it perfom, complex management algorithms and opacity of implementation do not make me feel very happy about flash.

[edit: typos]

This is the kind of thing that ISO900x systems could be quite good at dealing with.

Unfortunately, quality assurance systems are often just a paperwork exercise so a company can get a logo for their letter footings, so all those goods-in checks and route cards and batch numbering etc end up not being so much use.

It might feel like he's going to a lot of effort. He should be able to spec a part; get quotes; order; do minimal checking of the delivered product (does the pack say "4 GB MicroSD? does the device have printing that says "4 GB"?); and then use the part. The lengths he's going to are normally saved for when you're forced[1] to use obsolete stock from brokers - you don't have much idea about who has had those parts or how they've been stored[2] and so you want to put them through rigorous testing. (Especially because the situation means they're usually going into things that must not fail.)

[1] I was working on a device where the manf had specified a Bourns resister network at 1% tolerance. That design had been approved and certificated and was thus "locked in" - there was no possibility to change that device. 1% devices were not available. The manf offered a complicated alternative (testing a bunch of 5% devices) but PHBs declined to use that ("What? We buy stuff and we might not use it?") and thus we lost that bid.

[2] Certificates of conformity are easy, everyone offers them. But there's a further requirement for some manfs of "traceability" - each component must be traceable through your paperwork (and the paperwork of your suppliers) to the batch where it was made. In theory this is great. If you get a bunch of devices failing you can look at what device is failing, and see where else you've used that device, and predict other failures. In practice the paperwork is nonsense; or no-one is skilled enough to do that kind of analysis. (I guess the huge manfs might do it, but people building 30,000 off builds don't tend to have the time or software or expertise to do it.)

tl:dr QA for hardware is often a joke and there's plenty of room for disruption there.

This is pretty normal in the electronics industry. From major suppliers, in the 90s we had 74hc logic which was just relabelled 74ls and some east German clone z80s delivered as genuine zilog parts. You don't usually notice until something fails test or you get a high failure rate in the field.

Comedy moment was the 1M resistors which were actually zero ohm shorts.

If there is a dollar to be made...

I've always wondered if we paeons truly know what the heck is going on in our chips. I mean, the basis of trust is pretty large .. but there is really nothing we can do to ascertain whether there isn't a backdoor in every single chip package, and indeed this has been discovered to be a real situation by the Pentagon (Chinese backdoors in milspec chips), so .. I can hardly imagine that we can do much more than just plain trust, and get over it.

Consumer flash cards are an example of a market where the purchaser has very little information to go on. The raw user available storage number is pretty misleading as the block remapping system (or absence of one) is a key determinant of reliability.

The fact that SanDisk and other name-brand manufacturers make a practice of switching out the underlying parts without changing the part numbers make reviews pretty much useless.

I'm always amazed by the lengths forgers go in order to successfully deceive their victims. Specially when only takes the same amount of work, or a little, to achieve uniqueness. Poor bastards, they don't know what they're missing.