Journal

Notes from the back end.

Writing on advanced packaging, physical AI, yield economics and the engineering of trust — for people who would rather read an argument than a press release.

Read the latest

Featured · Yield economics

The most expensive scrap in manufacturing

Why back-end defects hurt differentlyA front-end defect loses a die. A back-end defect loses a module carrying a dozen of them, each already fully paid for.

There is a structural asymmetry in semiconductor manufacturing that the industry has been slow to price correctly. A defect in the front end destroys a die that has consumed lithography, deposition and etch. A defect in the back end destroys a package containing several such dies — plus the interposer, the substrate, the bonding, the molding and the accumulated handling — all of it already paid for.

On a 2.5D module with three HBM stacks and two compute dies, the cost of a late-stage failure is not a marginal yield statistic. It is the entire accumulated value of the most expensive silicon in the supply chain, discarded at the last step, for a void that formed three operations earlier.

This is why perception at the point of formation matters so much more in packaging than a naïve reading of defect-density economics would suggest. Detecting a void at final inspection prevents a bad module from shipping. Detecting the process drift that produces voids prevents the module from ever being built.

And it is why yield in advanced packaging cannot be treated as a downstream reporting problem. By the time a yield management system has correlated the excursion, the material is gone. The only intervention that changes the economics is one that closes the loop between perception and control — in seconds, not shifts.

Recent

Writing

Coming · Autonomy

Four gates between a model and a bonder

Envelope, twin, autonomy gate, audit — why each exists, and what happens when one is removed.

9 min read

Coming · Perception

Volumetric defect detection at line rate

What makes CT and X-ray perception hard in production, and why fab-edge inference is not a preference.

11 min read

Coming · Process

Hybrid bonding and the end of tolerance

Sub-micron alignment with zero particle tolerance changes what "process control" has to mean.

8 min read

Coming · Data

The correction is the product

Why a supervised engineer correction is worth more than a thousand unlabelled inspection images.

7 min read

Coming · Industry

Nobody owns the sequence

Every incumbent owns a tool. The decisions that matter live between them, and that is the whole opportunity.

10 min read

Coming · Craft

The engineers who cannot be hired

A structural shortage that no recruitment budget resolves, and what encoding judgement actually requires.

8 min read

Chipira is pre-launch; the journal is being written alongside the design-partner programme. Ask us for early drafts — we would rather have them read critically than published unchallenged.

Editorial policy

What we will and will not publish here

No customer data, ever, including anonymised process detail that a knowledgeable reader could attribute. No modelled figure presented as a measured one. No competitor comparison we would not make to their face.

What we will publish: arguments that could be wrong, positions we have changed, and the reasoning behind engineering decisions that a reader might reasonably dispute. If a post does not contain something we could be held to, it is not worth writing.

  • Modelled figures marked as modelled, always
  • No anonymised customer detail, however tempting
  • Positions we have reversed, stated as reversals
  • Corrections appended, not silently edited

The most useful thing a vendor ever sent me was a post explaining why their own approach might not work.

Process integration leadIDM back-end

Composite drawn from design-partner and industry conversations. Illustrative, not a customer endorsement.

Positions

Four things we believe that are not consensus

The back end is the constraint, not the front end

Lithography gets the attention and the capital. But the number of accelerators that ship this year is decided by bonding, stacking and molding yield — and by how quickly a new package can be ramped.

  • Value has moved to the module
  • Scrap cost is maximal at the last step

Perception alone is a smaller business than it looks

Better inspection tells you about failure sooner. It does not prevent it. The economics change only when perception is wired to control, which is a much harder product and a much larger one.

  • Detection is necessary, not sufficient
  • The loop is where the value is

No incumbent can build this

Not through lack of capability, but through structure. A vendor that sells a bonder cannot credibly optimise a line containing five competitors’ tools, and its customers know it.

  • Neutrality is a structural asset
  • It is also permanently unavailable to OEMs

Trust is an engineering problem, not a sales one

A fab does not grant write access because it likes you. It grants it because envelopes, twins, audit and revocation make the grant reversible and bounded. That is architecture, not relationship.

  • Bounded, reversible, audited
  • Earned per workflow, per measurement

Themes

What we write about

  1. 01 Yield economics

    Why a point of yield on a CoWoS-class package is worth more than most manufacturing software budgets.

  2. 02 Physical AI

    Perception, control and simulation applied to processes where the feedback is physical and the cost of error is material.

  3. 03 The engineering of trust

    Grounding, envelopes, twin validation, immutable audit and graduated autonomy as engineering surfaces rather than compliance artifacts.

  4. 04 Advanced packaging

    Hybrid bonding, HBM stacking, fan-out and panel-level, molding and underfill, warpage and the physics behind each failure mode.

  5. 05 The craft problem

    Encoding judgement that exists only in a few thousand heads, and what happens to an industry when those heads retire.

  6. 06 Building in the fab

    What it is actually like to sell, deploy and iterate software inside a semiconductor manufacturing organisation.

Context we keep returning to

The numbers behind the writing

  • $30B Advanced-packaging automation, process control, inspection, test and fab software.
  • ~16% Annual growth, pulled by AI accelerators, HBM and chiplets.
  • 12+ Tool classes on a line, none of which can see the others.
  • 1 Number of packaging engineers who can currently be in two places at once.

Elsewhere

Where else we turn up

Standards and industry bodies
Participation in packaging, warpage and defect classification standards conversations, because interoperability is our whole thesis.
Design-partner working sessions
Closed technical sessions with process, quality and OT teams — usually more useful than any conference talk.
Conferences
Advanced-packaging, reliability and physical-AI events where the audience is engineers rather than buyers.
Direct
The fastest way to get our thinking on a specific problem is to ask us about that problem.

Stay close

Argue with us before we publish.

We would rather send you a draft and have it dismantled by a process engineer than publish it and be right in private.