A fab's electrical supply is split into independent trains, sometimes two and often more, so that the building can lose one and keep running. That redundancy is real only if the tools respect it. Each tool has to be fed from one train, and each tool model has to be spread across all of them. Get either wrong and losing a single train takes down far more of the fab than it should.
What unbalanced looks like
The common failure is a tool with mixed trains: some of its loads fed from one train and some from another. That tool now depends on every train it touches, and losing any of them takes it down. A tool mixed across two trains is exposed to twice as many failures as a tool on one. Across a fab with mixed tools, a single train drop knocks out tools that should never have been affected.
The second failure sits one level up: every unit of a tool model on the same train. Say all of a fab's units of one etch model hang off Train A. Train A drops, and that etch step goes to zero. Every lot that needs it queues, wafers pile up upstream, the tools downstream run out of work, and a fab that is still mostly energised behaves as if it were down.
What balanced buys you
- One train per tool. Each tool is fed whole from a single train, so it is exposed to that train and nothing else.
- Every model across the trains. The units of each tool model are spread across all the trains. When one drops, every process step keeps running on the units fed from the others, at reduced capacity. The line slows down, but it keeps moving.
Balancing costs nothing in equipment: the panels, the cable, the switchgear are all the same. The only difference is which train each load is connected to, and that is a design decision. Make it well and you get resilience at no extra hardware cost. Make it badly, or not at all, and you've paid for multiple trains and still have single points of failure scattered across the tool fleet.
Multiple trains only buy redundancy if the tools respect them. Otherwise you have redundancy in the switchgear and single points of failure in the process flow.
Why most fab designs don't consider it
We've reviewed the electrical designs of a number of fabs, and in most of them train balance isn't even taken into consideration. That isn't carelessness. Done by hand, balancing is too complex to achieve, so it falls out of scope.
- Balance pulls against proximity. The natural move is to connect each load to its nearest panel. That's how a single tool ends up on mixed trains, and how a bay full of identical tools ends up on one. Fixing either means sending loads to panels on other trains without blowing up run lengths or panel capacity.
- It's a combinatorial problem. Every load of every tool, every tool model, every candidate panel, every train, all at once, with capacity limits on each panel. A change made to fix one model's spread can break another's. It doesn't decompose into problems a person can solve one at a time.
- The layout keeps moving. Tools are added, swapped and relocated throughout the build. A balance that was right at one design revision drifts with each change after it, and nobody re-checks thousands of assignments every time a tool moves.
So in practice loads get connected to whichever panel is nearest, and the train split is whatever falls out of that. Nobody chose it, and nobody checks it until a train drops.
Balance as a step in the run, not a review
In FacilityConnect, balancing is a stage of the Optimizer's auto-design run, not something checked afterwards. Once candidate panels are laid out for every train and every tool is connected, the run re-routes loads so each tool sits on a single train and every tool model is spread across the trains, while still holding panel capacity and keeping runs short. The redundancy is a property the run enforces, not one a reviewer has to find.
Because the run is cheap, it doesn't happen once. When the layout changes, the Optimizer re-runs against the living model, and the train balance is restored along with everything else. The balance at the last revision is the same as at the first, and every change is verified against it rather than assumed.
There is one honest limit. A tool model with a single unit can't be spread across trains. That's a question about sparing, not distribution. For every model with two or more units, the spread is the Optimizer's job, not a reviewer's.