The card that came with your dash cam: better than you think

Why test the card that came in the box?

The standard advice about bundled memory cards is to bin them. It is repeated confidently and almost never tested, which is exactly the kind of claim this lab exists to check. Bundled cards are also the single most-used cards in dash cams, because most people never change the one the camera came with. If that card is quietly bad, a lot of footage is at risk. If it is quietly good, a lot of people are being told to spend money they do not need to spend.

So we ran one. This card came out of a dash cam, it has no retail packaging, and everything we could claim about it up front came off the card face itself: Kioxia, Exceria G2, 128 GB, microSDXC, C10, U3, V30, UHS-I. That already makes it more interesting than the no-name the advice assumes, because Kioxia is the former Toshiba memory business, one of the handful of companies that actually manufactures NAND rather than buying and relabeling it.

Every figure below comes from one run on 2026-07-25, stored in a single results file, and quoted here exactly as recorded. Where a number needs a caveat, the caveat is next to it.

What Kioxia promises, and what the card did

There is no box, so the claims come from two places and we keep them separate. The card face prints its capacity and its class marks. The speed figures come from Kioxia's own published spec for the Exceria G2 128GB, which states 100 MB/s read and 50 MB/s write. Watch the write bar in particular: it does not stop at the dashed line where its own claim sits.

The full scorecard. Class floors are judged against the worst ten seconds of sustained writing, not the average, because a class rating is a promise the card has to keep for the whole recording.

Two of those rows deserve a note. The read figure looks like a miss at 97% of claim, but 96.8 MB/s is where the UHS-I bus itself runs out of road: in its fastest mode UHS-I tops out around 104 MB/s of raw bandwidth, and real ceilings land in the mid-to-high 90s once protocol overhead is paid. The card is as fast as the interface allows, and Kioxia's "up to 100" is effectively a statement about UHS-I rather than about the flash.

The A1 row is deliberately marked unprinted rather than met-as-claimed. Kioxia's retail spec for this model does list A1, but this card face does not print one, and a bundled unit may be an OEM version rather than the retail SKU. It clears both A1 floors in our test, so it earns the row. It just should not be scored against a promise this particular card may never have carried.

Why would Kioxia under-rate its own card?

A 50 MB/s write claim on a card that commits at 65.3 is not modesty. The printed number has to hold for every unit of every production revision for the life of the model, across NAND suppliers and controller respins, on the slowest host the card might ever meet. So the claim is a floor the manufacturer will defend, not a description of the card in your hand.

Kioxia say this themselves, more plainly than most. Their spec footnote reads: the figures are "the best values obtained in a specific test environment" and they "warrant neither read nor write speeds in individual devices." That is a manufacturer stating outright that the number on the sheet is a ceiling measured under favorable conditions, not a promise about your card. Ours happened to land above it, which is the same pattern the SanDisk High Endurance showed when it beat its own printed write claim by 2.24 times. Two cards, two brands, both quietly better than their labels.

The same logic explains V30 on a card that holds V60. V30 is what the dash cam and 4K camera market asks for, so it is the promise worth making, and headroom above it is margin the manufacturer keeps in reserve rather than in marketing. Which is why you cannot rank cards by their labels: two cards both printed V30 can differ by a factor of three in the number that decides whether your camera drops frames.

Why a 128 GB card only gives you 124 GB

Put this card in a computer and it reports 123.7 GB. The label says 128. That is 4.26 GB you paid for and cannot use, and it is the one genuine complaint this card earns.

The first thing to rule out is the boring explanation. It is not the dash cam's formatting: the physical disk reports 123.77 GB and the partition fills essentially all of it, so there is no hidden unallocated space to reclaim. It is also not the usual decimal-versus-binary confusion that makes every drive look smaller in Windows. This is the card telling the host, honestly, that 123.7 GB is all it has.

Kioxia do disclose it. Their spec page lists an approximate usable capacity of 115.2 GB for this model and explains that "capacity is based on installed flash memory and not user available memory as part of the memory is used for management functions." That figure is in binary units, which works out to 123.7 GB in the decimal units the label uses, and our card matched it to within 0.04%. So the card delivers exactly what its maker says it delivers, in a document approximately nobody reads before buying a dash cam.

Credit where it is due: publishing the number at all puts Kioxia ahead of manufacturers who do not. But disclosure is not the same as the label being accurate, and it is worth being clear about what the fine print does and does not settle.

Is the reserve a fair excuse?

Partly, and the honest version of the argument is more interesting than either "they lied" or "it's fine, they disclosed it."

Every card reserves space. The question is where the reserve comes from. A card sold as 128 GB is built from 128 GiB of raw flash, which is 137.4 GB in the decimal units the label uses, so roughly 7% is already invisible before anything is advertised. The SanDisk takes its entire reserve out of that raw headroom and still hands you the full 128 GB. This Kioxia takes about 10% of raw, which means part of its reserve comes out of the number printed on the front.

Both cards reserve. Only one of them makes the buyer pay for it. That is the fair criticism, and it does not require accusing anyone of lying: a competitor solved the same engineering problem without spending the customer's capacity.

It also raises a question with no good answer. Nothing sets a floor on how much a manufacturer may hold back from the advertised figure. If 4 GB is acceptable because a spec sheet mentions it, so is 20. The only thing standing between a buyer and that is convention, competitive embarrassment, and somebody actually measuring. Which is most of the reason this lab exists.

Where it is not a scandal: this card is nothing like the counterfeit we tested, which claimed 128 GB and exposed 34.3 GB, or 26.8% of its label. The Kioxia delivers 96.7% and every byte of it is real. Those two facts belong on the same scale, and they land nowhere near each other.

Five minutes flat out: the write curve

Consumer cards frequently write fast into a small cache and collapse when it fills. That is the failure mode that ambushes long recordings: fine for the first few gigabytes, then a cliff. So the sustained test writes flat out for five minutes with the cooling fan off, logging every second, while a class ladder keeps score of which speed ratings the card is still holding.

The test also starts honest. Before the clock runs, a settle pre-pass writes untimed until throughput stabilizes, so the measurement begins on a controller that has finished reacting to the random-write test before it. Without that, the verdict number would partly be measuring recovery rather than the card's real rate. Hover the line for the exact sample under your cursor.

Burst 63.5, steady state 63.4, worst ten seconds 62.6. The worst ten-second window sits 1.3% below the steady state, and there is no knee anywhere in the trace: the cache analysis returned NO_KNEE_STABLE, meaning the card never found a cliff to fall off. Every one of the 300 samples landed between 56.6 and 67.1 MB/s, a spread of about 15.6% peak to trough.

That spread deserves a word, because we published the same figure for the SanDisk High Endurance and it was about 8%, roughly half this one. Second to second, this card is the twitchier of the two. It does not change either verdict, and here is why: a class rating is set by the worst sustained window, not by the jitter around the mean, and 62.6 MB/s clears the V60 floor with room. A camera writes into a buffer that swallows a single slow second whole. What the wider spread does tell you is that the Kioxia's controller works harder to hold its line, which is worth knowing and is not worth worrying about.

That worst window sits at 1:00, the earliest moment a complete ten-second window exists to score, and it is 1.3% below the steady state. So it marks where measurement opens rather than a card digging out of a hole, which is why the endurance section can say this card needed no recovery without contradicting the chart.

Averages hide the thing that actually ruins footage, though. A card can post a fine mean and still freeze for two seconds, and two seconds is a hole in your recording no average will show. So the suite times every individual block it writes and keeps the tail. Across 18,141 blocks the slowest single write took 106.7 milliseconds, at second 287, on a file roll: the moment the test closes one file and opens the next. Nothing crossed 500 ms, nothing crossed a second. For a 4K30 dash cam writing roughly 6 MB/s into a buffer, a 107 ms hiccup is invisible.

Is the space real? Every byte, verified

A counterfeit card lies about its size. A controller reprogrammed to report more than it has will accept writes forever, silently wrapping new data over old once the real flash runs out. The only test that cannot be fooled is the brute-force one: fill every free byte with seeded pseudorandom data, then read all of it back and compare against what the seed says each byte must be. Every block carries a unique stamp, so a controller cannot fake capacity by recognizing repeated data.

This matters more than usual for a bundled card, because a camera that ships with a counterfeit is exactly the scenario nobody checks. It came back clean.

123.7 GB written across 116 files, 123.7 GB verified, zero mismatched bytes, zero read errors. The card holds exactly what it says it holds. The complaint from two sections ago stands separately: what it says it holds is 4.26 GB less than what the label advertises. Those are different claims and both are true.

The verify pass doubles as a measurement no manufacturer advertises and every user eventually cares about: how long it takes to get footage off. Reading the whole card back took 22 minutes 23 seconds at 92.2 MB/s wall-clock. Fill this card and budget that.

One number in the fill deserves the same treatment. It is quoted at 65.9 MB/s, which is the rate across 1,878 seconds of actual I/O, but the wall clock on that phase was 32 minutes 42 seconds, or 63.1 MB/s end to end. The 84-second difference is time spent between write operations rather than inside a slow one: file opens, allocation, the flush at each close. Worth stating because a speed quoted beside a duration should reconcile with it, and on our SanDisk run an unexplained pause made that gap more than twice as large.

The whole-card fill also settles the cache question at full scale. The same knee analysis that ran over the five-minute test ran over all 123.7 GB, and found nothing: it opened at 63.8 MB/s and closed at 63.9. A card written from the first byte to the last at one unbroken rate has nowhere left to hide a slow tier.

What loop recording actually does to a card

Here is the most revealing number in the run, and it is not a speed. Random 4K writes are what loop recording looks like to a card: event-file saves, filesystem updates, and above all the constant overwrite churn of a camera that never stops. A dash cam card lives in that workload permanently. So the bench runs three consecutive 20-second random-write passes specifically to see whether a card degrades as its flash translation layer gets churned. That layer is the firmware deciding where your data physically lands, and it is the part that gets tired first.

We put the lab's reference card through the identical test. Be clear about what that card is: an unbranded Class 4 32 GB card that writes sequentially at 11.4 MB/s, the slowest tier still sold and a bus generation behind this one. It is a floor, not a fair fight. Each bar is that card's percentage of its own best pass, so the shape of the collapse is comparable even though their absolute speeds differ by four times.

The reference card went 183, 76, 53 IOPS: a 71% collapse in sixty seconds of scattered writes, after which it needed nearly two minutes of continuous writing to climb back to normal. The Kioxia went 777, 782, 785. It did not fade at all.

Say that carefully, because the honest version is slightly narrower than the exciting one. A 1% rise across three passes is inside measurement noise, so the finding is no measurable fade, not it got faster. And treat it as indicative rather than proven: three 20-second passes measure how the controller behaves under churn, which is a good directional read on whether a card is built for loop recording, but it is not an endurance test and cannot tell you the card is still working in year three.

What it does say is narrow and useful. Under the workload that cost a floor-tier card 71% of its random-write speed, this one gave up nothing and needed no recovery. Its settle pre-pass reached a stable rate in 60 seconds or less, the earliest our analyzer can declare stability, immediately after the same abuse.

For comparison, the SanDisk High Endurance faded 2.1% in the same test. That card is explicitly sold on endurance and costs accordingly. The bundled Kioxia matched it on stability. It is slower in absolute terms, at 781 write IOPS against the SanDisk's 1030, so this is not a claim that the two cards are equals. What matches is the thing that decides whether a card degrades in a dash cam, and that is a genuinely surprising result for a card most people are told to throw away.

Two caveats stay in the open. In absolute terms the random figures are ordinary microSD numbers, 8.17 MB/s read and 3.20 MB/s write, which is why this is not a card to run an operating system from. And while it clears both A1 floors at 1996 and 781 IOPS, that is recorded as indicative rather than certified: the SD Association's A-class procedure differs from ours, and A2 is not testable on a USB reader at all because its floors assume command queuing that USB readers do not perform.

How hot does it get?

The sustained leg runs with the bench fan off, so the card heats the way it would inside a sealed camera body, and the thermal frame is captured the instant the write stops, because a card sheds heat in seconds. The question is not just the peak. It is whether the card slowed down on the way there.

Here is the frame itself, captured as the timed write ended. The spot marker reads 143.3 F, or 61.8 C, at the hottest point on the card.

That is a rise of 67 F, or 37.4 C, over the room, and the speed trace above it never moved: the card was writing at the same rate at peak temperature as it was cold. It also ran cooler than the SanDisk High Endurance, which peaked at 151.4 F in the same test. That is not a virtue so much as arithmetic, since the SanDisk was writing 88 MB/s to this card's 63, and controller heat scales with how hard the controller is working.

Now the dash cam arithmetic, done honestly in both directions. Taken at face value the rise looks alarming: a car parked in summer sun reaches 140 F in the cabin, and stacking a 37 C writing rise on that lands past the 85 C these cards are rated for. That is where a lazier review would stop, and it would be wrong. Our rise came from writing 63 MB/s continuously. A 4K30 dash cam writes about 6 MB/s, roughly a tenth of the data rate, so the bench is deliberately generating a thermal load no camera will impose.

The honest reading: this card's worst-case self-heating is real, a dash cam will not come near invoking it, and 61.8 C leaves comfortable headroom under the typical 85 C rating even at full abuse. What the number does tell you is that thermal headroom is not unlimited, so mount choice matters. A camera baking against black dash plastic in direct sun starts from a far worse ambient than a bench at 76 F.

What can the Kioxia Exceria G2 128GB record?

Every row is judged against the worst ten seconds of the run, 62.6 MB/s, not the average. Averages do not appear here on purpose: your camera does not record over the average, it records over the worst moment.

The V60 pass is genuinely close, clearing the 60 MB/s floor by 4%. That is still a pass, and it is two full grades above the V30 printed on the card, but it is worth being precise: a card that holds V60 by 1.04 times has less margin against that floor than the SanDisk's 1.47 times. V90 is marked not applicable rather than failed, because sustaining 90 MB/s requires the UHS-II bus and no UHS-I card can reach it.

One cross-check ties the run together. The cache-leak canary probe wrote at 63.6 MB/s. The calibration pass wrote 66.4. The sequential test averaged 65.3. The settle pre-pass stabilized at 63.3. The five-minute sustained test held 63.4. The whole-card fill wrote 123.7 GB at 65.9. Six independent measurements, three different engines, spanning 64 MB to 123.7 GB of data, all landing inside a 5% band. That agreement is the strongest evidence that the methodology is measuring the card rather than the tool.

Bottom line: the advice was wrong about this one

Keep it. On the numbers that decide whether a dash cam card survives, the bundled Kioxia behaves like a card you would have chosen on purpose: 65.3 MB/s committed against a 50 MB/s spec, a V60 floor against a printed V30, 62.6 MB/s at its worst moment in five minutes flat out, no cache cliff anywhere across 123.7 GB, no stall longer than 107 milliseconds, and no measurable fade under the churn that collapsed a floor-tier card by 71%.

The one real deduction is capacity. You paid for 128 GB and the card hands you 123.7, and while Kioxia disclose that in a spec sheet, the number on the card face is still 128. A SanDisk with the same label gave us the full amount, so the shortfall is a choice rather than a law of physics. It is worth about 19 minutes of footage.

The broader point is the one worth keeping. Two cards have now come through this bench and both beat their own printed write claims, one by 2.24 times and one by 1.31. Manufacturers publish floors they can defend for years across every production revision, not descriptions of the card in your hand. Which means a spec sheet cannot rank cards, and the only way to know what you have is to measure it.

This run is now the reference fingerprint for this card in the Card Lab. If a future Exceria G2 shows a cache knee, a random-write fade, or a different delivered capacity, the parts changed, and this page is what it gets compared against.