SD/MMC cards, a year later

[John Watlington]

Our experiment with SD/MMC cards as main storage continues.

One frustration has been the rate of change in the SD industry.
The dominant model from a vendor in a particular size and speed
may only be in production for three or four months before being
superseded.  Vendors are reluctant to properly inform Quanta of
changes which might require retesting.   The result is that SD card
certification is an ongoing process.

A disturbing trend has been the increased error rates and
decreased device lifetimes brought by higher density devices.
And these are occuring throughout the industry.

A batch of 2GB class 2 microSD cards obtained a year ago from a
particular manufacturer averaged around 10 TB written before
failing, with few transient errors.   A batch of 2GB class 2 microSD
cards from the same manufacturer today failed with more than
half corrupting their FS after only 1TB of writes.  The devices
wear out around 2-4 TB of writes.   I'm seeing the same error
distribution on 4GB parts from the same manufacturer, and
similar problems of early filesystem corruption from other
manufacturers as well.

Of the last five SD card models we've tested, we rejected three
of them for failing to survive 3 TB of writes.   In some size/speeds,
we only have a single vendor/model qualified --- always on the
brink of being end-of-life'd.

While the Armada 610 SOC being used in XO-1.75 does have a
"raw" NAND Flash interface, it's design faces the same problem
as the CaFE in XO-1 --- it will only be useful for today's (yesterday's)
NAND Flash parts.  Each new generation tends to introduce higher
demands on the (hardware implemented) error correction algorithms.
Nonetheless, we are going to try testing it with UbiFS and
considering whether any durability improvement justify the
increased price (now up to 2x) and built-in obsolescence of
using raw NAND chips.

On the positive side, the SD interfaces on XO-1.5 have been
exercised extensively and we have eliminated the bit errors that
plagued some XO-1 motherboards.

Overall, the industry is even more behind microSD than a year
ago, although eMMC (roughly the same protocol and all the same
automated wear levelling issues but 8 bits wide with multichip modules
soldered directly to the PCB) is gaining acceptance in the cell-phone
industry.

Cheers,
wad