What to Check on a Flash Sample Before You Run It in Production

TL;DR

A pre-production flash sample only earns its place if you test it the way the bulk run will behave. Use ISO 2859-1 acceptance sampling with General Inspection Level II as the default — sample sizes of n=125 for a 1,201–3,200 lot, n=200 for 3,201–10,000, n=315 for 10,001–35,000 — and set AQL by defect class: critical 0, major 2.5, minor 4.0. The one test that proves capacity is write-and-read-back, which fills the whole nominal space and reads every byte back; H2testw on Windows and f3write plus f3read on macOS and Linux do this, and a fake reports Data LOST past its real size. Inspect at incoming, in-process and outgoing stages, not just at the end. WINTEL builds TF cards from 4GB to 256GB and UDP modules from 128MB to 128GB; approve the sample against this checklist before the line runs.

A sample is a promise, not a courtesy

The pre-production sample lands on your desk as a small, clean object. It is easy to glance at it, nod, and release the bulk order. That glance is the most expensive minute in the programme, because the sample is the only thing you personally verified before committing to thousands of units. If it lies, the line repeats the lie.

Do not approve a sample on appearance. Approve it on the tests that the bulk run will fail if the sample was wrong. The sample is a promise about every box that follows; treat it as a contract, not a formality.

Size the inspection to the lot, not to habit

Acceptance sampling under ISO 2859-1 gives you the sample size from the lot, with General Inspection Level II as the default when a contract is silent. For a lot of 1,201–3,200 units the sample size is n=125; for 3,201–10,000 it is n=200; for 10,001–35,000 it is n=315. The number follows the batch, not a round habit of “check ten.”

Inspecting too few hides a bad lot; inspecting by feel hides your own method. Pin the level and the size in the PO so the supplier knows the bar before they ship the sample, not after you dispute it. A sample count you cannot defend is a sample count that will not hold.

AQL is a plan, not a permission

The AQL values are the part buyers misread. Common assignments by defect class are critical 0, major 2.5, minor 4.0. Those are planning parameters, not a defect allowance — AQL 2.5 does not mean 2.5% bad is fine. It sets the acceptance number for the sample: at sample size 125, AQL 1.0 accepts Ac=3 and rejects Re=4, while AQL 2.5 accepts Ac=7 and rejects Re=8.

Set critical at zero for anything that breaks safety or data. Major covers function; minor covers cosmetics. Write the three numbers in the order so “a few defects” can never become a negotiation after the fact. The AQL you did not set is the AQL the supplier will choose for you.

Prove capacity the only way that works

The single test that does not take the device at its word is write-and-read-back. You fill the whole nominal capacity, then read every byte back and compare. A card that lies about its size cannot fake this: it runs out of real cells and the read-back shows the gap. Tools such as H2testw on Windows and f3write plus f3read on macOS and Linux implement exactly this, and a fake reports Data LOST, Corrupted or Overwritten past its real capacity.

A reader check that just lists the free space does not prove anything — the device reports what it claims, not what it holds. Skip write-and-read-back and you have verified the label, not the memory. For a bulk order this is the one test worth running on every sample, because capacity fraud is the failure that scales with the lot.

Split the inspection across the process

One final check is too late to catch most problems. Inspection is normally split into incoming, in-process and outgoing stages rather than one gate at the end. A sample you approve should represent all three, because a part that passes only at the finish can still hide a process drift the bulk run will expose.

Ask the supplier what they check at incoming, what they check mid-line, and what they check at pack. A sample pulled only from the outgoing tray tells you about the tray, not the process. The line you are about to trust runs all three stages every day, so your sample should speak for all three.

What to write on the sample approval

Five items belong on the approval you sign. First, the sample size and inspection level from ISO 2859-1. Second, the AQL by defect class, critical at zero. Third, a write-and-read-back capacity result with the tool named. Fourth, the read and write speeds measured, not quoted. Fifth, the die grade and maker behind the part, because the sample’s behaviour depends on it.

A signature without those five is a wish. The supplier who hands you a sample and asks you to “just confirm it looks good” is asking you to skip the only checks that matter.

Where WINTEL fits the sample stage

WINTEL supplies TF cards from 4GB to 256GB at 15 × 11 × 1 mm and 0.25 g, and UDP modules from 128MB to 128GB at 11.3 × 24.8 × 1.45 mm and 0.819 g, with read/write speed specified per order. The sample we send is built on the same line and the same specification as the bulk run, and we document the inspection stages so the sample you approve is the one the line will repeat.

Tell us your lot size and your AQL by defect class and we will run the sample against ISO 2859-1 with write-and-read-back on capacity, so the approval you sign is backed by the test, not by the photograph.

The sample you skip is the risk you ship

The bulk run is a copy of the sample, good or bad. A sample you approved on looks becomes a thousand units you approved on looks. A sample you tested with the right size, the right AQL and a real write-and-read-back becomes a line you can trust.

The checklist is not bureaucracy. It is the difference between a sample that reassured you and a sample that proved itself.

What to do when the sample fails

A failed sample is cheaper than a failed line, and that is the whole point of testing it first. If write-and-read-back shows Data LOST past the nominal capacity, stop the bulk order and send the result back with the tool name and the lot size attached. If the AQL math rejects the sample, the contract already tells you the next step, so use it rather than negotiating the failure by feel.

The sample that fails on your bench spares you the batch that fails in your customer’s hand. Treat the failure as the saving it is, document it, and re-qualify the replacement the same way — the second sample earns approval only by passing the same tests, not by arriving later. A sample you excused is a risk you shipped, and the bulk run is only ever a copy of the sample you approved.

---

If you want a pre-production sample tested the way the bulk run will run, send us your lot size and your AQL by defect class. We will run ISO 2859-1 sampling with write-and-read-back on capacity before the line starts.

Reach the team at cathy@go-wintel.com or open the request a quote form.