How we test and score memory cards

Every byte committed, every byte verified

Two rules shape everything else. The first is that a write does not count until the card confirms it. All test writes use unbuffered write-through handles and the timer stops only after the flush returns, so nothing is credited to the card that is actually sitting in the computer's memory. This is why our write speeds read lower than burst benchmarks like CrystalDiskMark: a burst benchmark measures a card's small RAM buffer for a few seconds, and we measure what the card sustains when the data has to land. Read speeds match CrystalDiskMark within a couple of percent, which is how you can tell the gap is method rather than a slow sample.

The second rule is that data is checked, not assumed. Every test byte is generated from a seeded random stream and compared on read-back, so a card that drops, zeroes or repeats data cannot hide it. A cache-leak canary runs before anything else and aborts the whole run if any read comes from the operating system instead of the card.

The tests only ever create and delete their own files, and the tool refuses to touch a fixed disk without an explicit override.

What actually gets measured

The verdict that decides a video speed class always comes from the worst ten-second window, never the average. A class rating is a promise the card has to keep for the whole recording, and your camera does not get to record over the average.

How the score is built

Every review carries a score out of 5. It is the mean of five dimensions, each scored out of 5 in half-point steps, with the mean rounded to the nearest half point. Since September 2026 every dimension is read off a fixed table, published below: each table has one measured input, its edges sit where the measurements on this bench actually fall, and an input exactly on an edge takes the higher band. The tables are the same for every card. Whether a number is enough for the job a card is sold for is a question the verdict answers in words, and the picker tools answer per device; it is not a reason to move a row. Every review publishes its own breakdown with the inputs behind it, so you can see which dimension moved the number and disagree with us specifically rather than generally.

How each dimension is scored

Each table has one input, taken from the card's results file the same way every time, so a score can be recomputed by anyone with the numbers. Every edge reads "or more": an input exactly on an edge takes the higher band. Half points are the smallest step. The published breakdown in every review carries the inputs its rows were read from.

Claim accuracy

Every claim printed on the card face or its retail packaging is measured, and the ratio of measured to claimed is taken for each: a printed write speed against the three-pass committed sequential write, a printed read speed against the sequential read on a reader proven capable of exceeding the claim, and every class mark against the worst ten seconds of sustained writing divided by that class's floor. The lowest ratio sets the score. A read claim above the ceiling of every reader on this bench is marked not settled and left out, because a number the bench cannot test is not evidence against the card. App class marks are not testable on a USB reader, endurance hours are trusted rather than tested, and a manufacturer specification that is not printed is not a claim. Capacity enters only as a gate: below the line in the capacity table, this row is 1 as well. A look-alike or wrong SD family mark costs one point.

The step from 4 to 3 at 85% is deliberate. A printed number missed by a few percent is a rounding; one missed by more than 15% is a false claim, and the three write claims this bench has caught short sit at 83.4%, 79.5% and 71.3%.

Sustained performance

The input is the worst ten-second average across the five-minute uncooled sustained write, started on a settled controller. If the whole-card fill finds a cache cliff the five-minute test did not, the post-cliff figure is used instead. A card that stops accepting writes for a second or more, reproduced on a second reader, scores 2 here whatever class it holds, and the instantaneous-floor rule below caps its overall score as well.

The top band starts at 75 rather than at the V90 floor on purpose. A card on the standard UHS-I bus cannot reach V90 no matter how good it is, so a 5 that required V90 would be a bus-tier gate rather than a quality bar. Seventy-five is the midpoint of the V60 band, and three cards on the plain UHS-I bus clear it.

Endurance behavior

Two inputs. The first is fade: how much of its random-write speed the card gives up between the first and third of three twenty-second scattered-write passes. Anything under 5% is inside the bench's own settle tolerance and counts as no fade; a card that got faster scores as zero. The second is recovery: how many seconds of continuous writing after that churn it takes the card to get back to 90% of its steady sequential rate, which is where a dash cam's event-file churn shows up.

Then the recovery deduction, never below 1: back to speed within 20 seconds, nothing; 21 to 60 seconds, half a point; 61 to 120 seconds, one point; 121 to 300 seconds, a point and a half; not back to speed inside the run, two points. Every genuine card we have measured sits between no fade and 5.3%, with one at 16%; the reference card that collapsed to a third of its speed sits at 71%.

Capacity integrity

The card's declared user area, the size the card reports to the reader and the size your computer shows for the disk, is divided by the label in decimal bytes, taken to two decimal places, and scored from this table. It is the declared area rather than the formatted volume, so the filesystem a card ships with cannot move it. And it counts only once the whole-card fill has verified clean: every declared byte written, read back and matched, with zero mismatches and zero read errors.

A fill that fails to verify is re-run on the second card reader. If the failure follows the card, the card scores 1 and the disqualifier applies; if it stays with the reader, it is recorded as an anomaly and the clean run is scored. The shortfall inside the honest range is scored here and nowhere else: a label's own footnote says actual storage is less, so a card delivering 97% of it has not contradicted its packaging, and claim accuracy does not deduct for the same gigabytes a second time. Every genuine memory card we have measured declares between 96.7% and 100.1% of its label, in a few clusters this table separates, and every fake has failed to verify with real capacity below 52% of the label behind a declaration that reads higher.

Thermal behavior

The input is the headroom, in degrees C, between the card's peak surface temperature at the end of the uncooled five-minute write and the maximum operating temperature its maker publishes, 85 C for most cards and stated as assumed where the maker prints nothing. The peak is the hottest card-surface reading we have for the unit, normalized to the 76 F the bench is held at. A full-size SD card disappears inside the reader, so it is pulled the instant the write stops and shot face-on; the reading through the slot is never scored, because it under-reads by 10 to 14 F.

Then the throttle deduction, never below 1: a card whose write rate at the end of the uncooled five minutes is under 97% of its rate at the start loses a point, under 90% two points. No card on this bench has triggered it yet; the rule exists so the first one that does is scored the same way as the next. Heat per unit of work, the temperature rise per MB/s written, is published as a rank among the cards we have measured and is not scored: it rewards a fast card for being fast, which the sustained row already does, and it cannot see a card running above its own rating.

Adopted September 2026. The capacity table went first, on 15 September; the other four followed the same day, after every published score had been re-derived from its measurements. Until then those four rows were judgments against the cards that came before, and they drifted: the same 98.8% of label scored 4.5 on one card and 4 on another, and a card was marked down for printing no speed figure while another was not. Six published reviews carry overall scores these tables set differently, every one of them higher, because their rows were docked for things no measurement reproduces: the Botslab 128GB (3 to 4.5), the Newegg TeamGroup 128GB (2.5 to 3.5), the VIOFO High Endurance 32GB (3.5 to 4) and 128GB (3 to 3.5), the Samsung PRO Ultimate (4.5 to 5) and the SanDisk MAX Endurance (4.5 to 5). The Botslab, the Newegg TeamGroup and the VIOFO pair were amended on 19 September 2026, each with a dated note at the top saying what moved and why; the PRO Ultimate and the MAX Endurance follow the same way. Other reviews carry individual rows the tables set differently without changing their overall score, and each is brought onto the tables when it is next revised (the SanDisk High Endurance 512GB, the Vantrue Endurance PRO, the SanDisk Ultra 128GB and the FitcamX 64GB on 19 September 2026). The full derivation, every candidate table and every card's placement, is in the bench protocol.

Conventions we hold to

What we do not test, and why

A method is only credible if it names its own gaps. These are ours.

Who pays for the cards

We buy the cards we test, or they arrive inside a device we bought. One card in the corpus was supplied by its manufacturer: Vantrue sent an Endurance PRO 128GB as part of a review kit in September 2026, and that unit’s review says so on its own page. No manufacturer has seen a result before publication or had any editorial input, on that card or on any other, and each review states how that specific unit was acquired. Where a review links to a product it is usually an affiliate link, disclosed on the page, and the link never changes a number. CK Tech Check runs no advertising of any kind.

Updated September 2026. This paragraph previously read “no manufacturer has supplied a card”, which was true until the unit above arrived. Rather than quietly drop the sentence we have narrowed it, because the part that actually protects you is the rest of it: nobody outside CK Tech Check sees a number before you do. A supplied card is tested on the same bench, to the same protocol, and scored by the same rubric as one we paid for.

If a card we recommended turns out to be wrong, the correction goes on the page rather than quietly into the archive. Two of the reviews currently published carry amendments made after the fact, both marked.