Lab-tested storage, starting with memory cards

Memory cards: every one we have tested

Same bench, same reader, same suite, with one exception the table states outright: a microSD Express card carries PCIe on its second row of pins and cannot reach that bus through any ordinary reader, so it was run on both and appears twice, once per bus. Every other row shares one measurement chain. Write speed is committed sequential, which reads lower than burst benchmarks because every byte had to land before the clock stopped. Capacity is what the card actually delivered against the number on its label, verified byte by byte.

Six patterns across fifteen cards

Fifteen cards, one measurement chain. Read down the columns and the same shapes keep coming back, and not one of them is visible from a single review.

Where each card sits against the rest

A review can tell you a card held 88.0 MB/s. It cannot tell you whether that is good. Every ranking below is computed at build time from the same results files the table above prints, across twenty-seven cards and twenty-eight measured rows, and it is worth being blunt about what that makes it: these are percentiles of our bench, not of the market. We have tested these cards. We have not tested the ones we have never held, so n is printed beside every ranking on the page and you can weigh it yourself.

The rows are grouped by host bus and ranked only against their own group. Of the three buses represented here, UHS-I is the only one carrying enough cards (twenty-five) to support a ranking at all. Any bus below that line is listed with its measurements and no percentile, and earns one once it reaches eight cards.

Read the full autopsies

The bench tests its own gear too

CKBench is not only for review samples. Before a piece of storage earns a job on this bench, it gets the same suite as everything else, and the first one through was a USB flash drive about to become the boot device for our card-identity rig. It is not a memory card, so it is scored on its own terms and stays out of the card table above. It also produced the strangest finding on this page: a drive whose read speed depends on how the data on it was written.

What buyers actually ask

How these numbers are produced

Every figure on this page comes from one results file per card, written by a bench built for this: writes are committed and confirmed before the timer stops, a cache-leak canary aborts the run if the operating system serves a cached read, and every test byte is seeded so it can be verified on the way back. Cards are scored on five dimensions against the job they are sold for, and a card that misrepresents its capacity is capped at 1 out of 5 no matter how fast it is.

The full method, the rubric, the unit conventions and an honest list of what we deliberately do not test are all on one page. Alongside it sits the reference we built while identifying these cards: every SD manufacturer ID we have decoded, and an honest account of what that number can and cannot prove.