An inference pool on the fab's own network

A fab is the most vendor-dense floor there is, and the most protective of what happens on it. Overhead transport vehicles, wafer-handling robots, and inspection stations each arrive from a different tool vendor, each with its own vision and its own expectation of sending images outward to be analyzed. The inspection models that sit beside a tool have to answer inside that tool's cycle, and they have to do it on a network that, by design, has no path out of the fab. A pool on the fab's own network, in the fab's own rack, is the only place those two requirements meet.

What is on the floor

Overhead transport carries lots along the ceiling from tool to tool and lowers them to load ports, watching for a mis-seated carrier or a person beneath the track. Wafer-handling robots at each tool move wafers between the carrier and the chamber. Inspection stations capture images of wafers between steps and judge them against what a good wafer looks like at that step. Cameras on and around the tools watch for particles, for a door left open, for an operator where one should not be. Nearly every one of these came from a different supplier, and nearly every one does its vision on an island.

The images are the point. A wafer image, the recipe that produced it, and the yield that followed are the most sensitive material the fab owns. Each vendor's vision system was built to send that material somewhere to improve a model the vendor keeps.

Why the cloud and a GPU on every tool both fail

Each tool vendor's cloud is a separate uplink from a network that should have none. A fab segment is isolated for good reasons, and every vendor that asks for a path out is asking the fab to make an exception for a model the fab does not hold.

The cloud also fails on time. Inspection has to answer inside the tool's cycle, not a round trip away. A station that waits on a distant model holds the tool behind it and the lot in front of it.

A GPU on every tool solves the latency and nothing else. Inside a cleanroom, every piece of equipment is a source of heat and particles and a reason for a gowned maintenance visit. Compute scattered across tools is idle between cycles, serviced in the hardest room in the building to service anything, and still an island. A model learned on one tool is relearned on the next, because every vendor keeps its own.

What the pool serves

Czar Grid is several Czar nodes joined into one inference pool on the fab's segment. Overhead transport vehicles, wafer-handling robots, and inspection stations from many vendors are served from that one pool, over the fab's own network. Each vendor's software calls the pool within its own latency budget, and the fab's automation team sees every call on one record rather than on several vendors' partial logs.

Wafer vision stays in the fab. Inspection models run beside the tool, on the pool, inside the tool's cycle, and the images they judge, the recipes they are judged against, and the yield data that follows never leave the cleanroom network. Each judgment is written to the record with the model version that made it, for the fab's own review.

The fab learns on its own lots. Overnight, Adaptation improves the models on the fab's own captures, on the pool, with nothing leaving the segment. A defect signature first seen on one inspection station becomes part of the model that every station asks for, once Evaluation has checked it against the fab's own held-out images and the fab's own review has agreed to serve it. The fab stops relearning its own process once per vendor. Safety, stops, and motion limits stay on the robot as the vendor built them; the pool answers questions, it does not drive the machine.

What stays on site and who holds the record

Everything the pool holds is in the fab's rack. Substrate runs on the Grid: Inference serves the models, Retrieval and Memory hold what the fab has chosen to give the pool about its tools, steps, and known defects, Adaptation and Evaluation do the overnight work, Watchman verifies the integrity of what is running on every node, and Local Bridge permits only the routes the fab declares. On an air-gapped segment there are no routes out, and signed updates are carried into the rack on media.

Tasks that cross tool vendors, a particle sweep of a bay, a check of every load port on a line after a transport fault, a review of the night's inspection exceptions, are authored in Agency Studio on the Grid, bounded to what they may read, approved by a named role at the fab, and recorded so that a run can be replayed. The record is on the fab's hardware and is read by the fab.

How it is placed and bought

The Czar nodes go in the fab's own rack, in the controls room or the equipment space on the fab's segment, in a conditioned and secured room the fab already controls, and the fab's network carries inference into the cleanroom to the tools. The fab holds one subscription for the configured Grid, hardware, platform license, and sustainment in one line, fixed for the term, and adding a tool or a vendor to the pool does not move it. Where a tool vendor or integrator prefers to deliver the pool under its own name with its equipment, it does so as Czar Chassis. It is the same Grid, held by the vendor.

A fab already keeps its process to itself. The pool lets its vision work the same way: every vendor's tool served from one place, every judgment on one record, and every image still inside the segment where it was taken.