Blog · operations

Hardware version control: the half that lives at the supplier

Hardware version control covers the design files, so why does the wrong revision still reach the factory, and what has to change at the purchase order?

ChinaPCB Editorial 1685 words · 8 min read

The switching signals in this article

  1. 01 Your repository ends at the door Git, SVN or a PLM system controls what you approved. It has no field for the drill program, the stencil aperture or the test fixture the supplier builds from your data.
  2. 02 Ten artefacts, three controlled A revision has to travel through ten artefacts before a panel is cut. Three of them are yours. The other seven are created downstream from your package.
  3. 03 The word means a certificate Across 131 profiles the word revision appears 15 times, and every one is a certificate revision, a standard revision or a document date, never a board.
  4. 04 Electrical test forgives a mix-up A board built from a previous drill file is electrically consistent with itself. The netlist test passes, and the error stays dimensional until someone measures.

A republished Altium article on hardware version control is a fair account of the upstream half of the discipline. It covers Hardware Configuration Management as a set of baselines, formal change control and configuration audits; a Hardware DevOps loop in which DRC, ERC and BOM validation run as gates that can fail a release; the three occasions when a project should be cloned, including the one about a part reaching end-of-life; branching practice for hardware and firmware teams working off a stable main line; and the component library treated as a versioned artefact rather than a folder that drifts between engineers.

One line in that article is the whole downstream problem, and it is offered as a safeguard: clone the finalised project before releasing it to the fabricator, so that a clean pre-release copy is retained. That protects your copy of the design. It does not protect the shipment. Between your released revision and the panel that arrives, the revision identity is destroyed and rebuilt by people who never see your branch structure.

Ten artefacts, three of them yours

A revision has to survive ten carriers before a board exists. Counting them changes the picture of what version control is for.

CarrierWho holds itWhat a lag looks like
Design data and libraryYour repositoryThe one artefact you already control.
Release packageYour repository, then copiedTwo packages with similar names and different contents.
Purchase order lineYour ERP"Same as last time." No revision, no date, no hash.
Fabrication drawingSupplier, sometimes yoursRevision letter missing, or carried over from the previous build.
File names in the setSupplier's CAM deskRevision in one file name, a date in another, nothing in the drill file.
CAM job and drill programSupplierTooling generated from the previous package and never regenerated.
StencilAssembly houseAperture cut for revision A, still in the rack for revision B.
Test program and fixtureAssembly or test houseNetlist from revision B, fixture wired for revision A.
As-built recordSupplierA lot number with no revision beside it.
Board markingSupplier silkscreenNo revision on the board at all, so the field cannot be sorted later.

Three of the ten are in your repository, and the third is only there if someone makes the purchase order carry the revision. The remaining seven are created downstream. A branch strategy does not reach them, a tag does not reach them, and a release note does not reach them unless the note is quoted on the order.

The word “revision” means a certificate here

A buyer who asks a Chinese factory which revision it built is using a word the public record has almost no slot for. Across the 131 published profiles, the word revision appears 15 times in 13 profiles. All 15 refer to something other than a board:

  • Withdrawn certificate revisions. Delton in Huangshi still shows ISO 27001:2013 on its certificate page; Leuchtek in Zhejiang quotes ISO 9001:2008 and ISO 14001:2004, both replaced by the 2015 editions; Zonbo in Yuyao still prints ISO 9001 at the 2008 revision next to a claim of the 2016 automotive standard.
  • A standard revision printed wrongly. Compeq Huizhou lists IATF 16949 as the 2024 revision, which does not exist; the automotive standard in force is the 2016 edition.
  • A document revision date. The same sentence appears in several profiles: ask for the process capability sheet with a revision date. JDB and Yidong both use it, and it is the right instruction.
  • A market description. Stariver Huangshi is the only profile where a customer’s revisions appear at all, as “mid-volume orders with frequent revisions”. That is a description of demand, not of a control.

The vocabulary of configuration management exists in this industry. It points at the documents that certify the factory, not at the object you buy. That is why the question “which revision did you build” so often produces a certificate rather than a lot record: the answer has to be requested in a form the factory does not normally produce.

What the counts say about the carriers

The same 1,252 capability lines that describe one manufacturability review describe almost no revision control.

Element of a revision chainCapability lines that describe it
ECO, ECN or PCN handling0
First-article inspection0
Tooling or fixture control0
As-built, serial or date-code record0
Revision control as a service0
AOI or X-ray inspection2
Electrical test3
Traceability, inside a bundle of promises1

The single traceability row is the same one that carries the DFM promise: Yuejie’s “99% sample on-time delivery, 24-hour quick-turn, 2-hour quotation, free DFM and traceable manufacture”. Five promises in a cell, one of which is a revision control claim without a method attached.

The zeroes matter because of where tooling sits. A drill program, a rout fixture, an ICT fixture and an AOI program are all built from one revision. None of them is regenerated automatically when a new package arrives; someone has to decide to rebuild it, and the decision is normally made by looking at the purchase order. That is the carrier with no record and the highest blast radius, and it is invisible in every profile in this directory.

The stencil deserves the same treatment on the assembly side. Ask an assembly house for the stencil’s aperture revision rather than the stencil’s manufacturing date. The date tells you when the steel was cut. The revision tells you which pads it will paste.

Three ways a revision is lost

FailureSymptomWhen it surfaces
Mixed data setGerber at revision C, drill file at revision B. Holes missing, undersized or offset in one pattern.At first article if anyone measures; otherwise at the customer.
Tooling lagCorrect files, previous drill program, previous test fixture. A consistent offset across a whole lot.When the lot fails test as a block, or when the fixture will not seat.
Effectivity gapCut-in by stock level, so one order ships two revisions with no record of which is which.In the field, months later, during a failure investigation.

The first one is worth dwelling on, because it defeats the control most teams rely on. Electrical test compares a board with the netlist you supplied. A board built from a previous drill file against a current Gerber set is electrically consistent with itself: the copper connects where the netlist says it connects. The test passes. The error is dimensional, and dimensions are not what the netlist measures.

The second one is the reason to ask about tooling explicitly. A factory can hold your current package, route it correctly, and still drill it on a program generated for the previous build, because nothing in the workflow requires the program to be regenerated when the file name changes.

A purchase order that carries a revision

The purchase order is the only document that is present at every step and is written by the party that owns the design. Six fields do most of the work.

FieldWhat to write
Revision on the lineThe revision letter and the date of the package, not "latest" and not "as previously supplied".
One package, one hashA single archive with a checksum in the order notes, and file names that carry the same revision letter.
Effectivity"Cut in at lot N" or "from this order onward". If stock-level cut-in is unavoidable, name the lots on both sides of it.
As-built recordOne line per lot: the revision built, the date, the CAM job reference. Kept with the traveler, supplied with the shipment.
Board markingA revision letter on the silkscreen for changes to copper, drill or stack-up. Ask first whether the marking is allowed on that board.
First-article scopeWhat changed, not "full FAI". A component change is verified on an assembly; a stackup change on a bare board.

Add one rule to the commercial side: any supplier-proposed adjustment comes back as a written waiver tied to that revision. A verbal “we adjusted the pad” is how a delivered board stops matching its drawing, and it is the most common form of unapproved change in the whole chain.

Where the discipline is not the constraint

For a single prototype build, the six fields above are enough. A configuration board, a baseline audit and a formal change procedure would cost more than the board. The arithmetic changes at four points: repeat production, a second site, a regulated or safety-critical part, and any programme where two revisions must exist in the field at once.

For the regulated case, the sector certificates are a gate rather than a record. Aerospace and defence work expects configuration management inside the AS9100D scope, and automotive work expects it inside IATF 16949, but both are site certificates and neither gives you the lot record you need to answer a question about a board built four years ago.

The component library is the one artefact this discipline shares with the factory, and it is worth being precise about it. A footprint is a claim about a package, and the fab builds to your land pattern. A version-controlled library that is wrong is still wrong, which is why the components topic treats the library as a set of assertions about other people’s parts rather than as your intellectual property.

The two sentences a programme needs

Your repository answers one sentence: this is what we approved. The supplier’s records answer another: this is what we built. A programme is controlled when the two sentences agree, and by default only the first one is written down.

The fix is not more tooling on the design side. It is one line on the order, one hash on the package, and one record per lot — plus the willingness to ask a factory a question it has not been asked before. The fabrication topic covers the rest of the handoff, the testing topic covers what a test program can and cannot prove, and the verification guide is the check of who is actually receiving the revision.

Questions this article answers

Is a design-tool ECO the same as a manufacturing ECO?

No. A design-tool ECO synchronises the schematic and the layout inside one database. A manufacturing ECO controls the authorised revision, the effective date, the affected production data, the material disposition and the verification. Your purchase order references the second one, and the first one does not produce it.

Can I rely on the supplier building whatever is in the shared portal?

A portal link is a location, not an identity. The package behind it can be replaced without telling you, and the link cannot prove which revision was used. Send a package with a hash and name the revision on the purchase order line, then ask for the as-built record that ties the lot to it.

How do I confirm which revision was actually built?

Ask for one as-built record per lot naming the revision, and for a revision letter on the board where the change touches copper, drill or stack-up. Then verify it on first article against the changed feature. Zero of 131 published profiles mention an as-built or serialised record, so this is a written request rather than something you can read off a listing.

Does ISO 9001 or IATF 16949 guarantee revision control at the factory?

Both require configuration management inside the certified scope, which usually means the site and the product families named on the certificate. They do not give you a document you can read without asking, and no capability line in the directory describes ECO handling or first-article procedure. A certificate is a gate; the record is a separate request.

What is the cheapest control that prevents mixed data?

One package, one hash, one purchase order line naming the revision and the date of the package. Add a first article scoped to what changed. That set costs a few minutes of administration and catches the failure that electrical test cannot see.