The Best Kind of Failure

The latest and greatest computer design has a problem or a potential

Em Thompson had spent eleven years learning how a wafer fails. She knew the usual signatures: a particle in the resist, a dose error at the edge of the field, a via that never opened. What she did not know, on the third night of the shrink, was why identical segments were dying on chips that had never been powered at the same time.

The program was called Halve. The foundry had qualified a new lithography stack that could print the current chipset at half the linear scale, two times smaller on a side, four times denser on the die. Marketing wanted the headline. Reliability wanted the old margins. Em wanted the data. She sat in the burn-in lab with the racks humming and watched the failure map update in blocks of red.

"Same column," said Luis from the next bench. "Row forty-one through forty-eight. Three boards. They were not in the same oven."

Em pulled the lot genealogy. The three dies had come from neighboring reticle shots on the same wafer, printed in the same pass, within a few millimeters of one another. Their layouts matched down to the last dummy fill. Under the old process, that would have meant nothing. Under Halve, the metal pitch had dropped far enough that the coupling between adjacent lines sat inside the noise floor of the old models. She had signed off on those models herself.

They isolated a fourth die from the same neighborhood and left it dark. No power. No clock. The package sat in a shielded box at the end of the bench while the other three ran a standard memory march. Twenty minutes in, the dark die's thermal camera showed a hot stripe along the same column. Em killed the test. When she opened the box, the stripe had fused. The unpowered chip had burned a segment it had never been asked to use.

They ran it again with controls. Random dies from other wafers stayed cold. Dies from the same shot, even across different boards and different rooms, did not. If two of them toggled the same address range in the same window, a third from the set took the heat. The current did not travel on a wire between the packages. There was no wire. The active pair drove the idle twin hard enough to melt the matching segment, as if the pattern on one die had a claim on the pattern of the others.

"Process variation should have killed this," Luis said.

"Process variation is what we removed," Em said. "We printed them too alike."

Management called it a fatal defect. The shrink had produced a population of chips so close in geometry that a subset could be driven from outside their own package. A customer who bought three unlucky parts might see one fail because the other two had been busy. Yield models collapsed. The recommendation on Friday morning was to scrap the engineering lot, seal the wafers, and bury the mechanism in a reliability report no one outside the building would read.

The scrap order never reached the floor.

Two people in plain badges arrived before noon with a letter on agency letterhead and a technical annex that already contained Em's failure maps. They did not explain how they had the maps. They explained what they wanted. The lot would not be destroyed. The twins, as the annex already called them, would be sorted, serialized, and moved. A full production run of the Halve process would be purchased under a cover contract, not for performance, but for match. Hundreds of thousands of dies, binned by wafer neighborhood, held as sets.

Em was offered a chair in a room without windows and a choice that was not much of a choice. She took the chair.

The mechanism, once they stopped calling it a bug, was simple to describe and hard to believe. A twin did not need a radio. It needed a match. Drive two chips from the same set in concert, and the third, wherever it sat, dissipated the energy in the corresponding segment. The idle chip was not reading a packet. It was being written by force. Partial destruction was the channel. A melted line was a one. An intact line was a zero. The host chip paid for the message in metal.

Three chips made a closed set. One stayed with the sender. One stayed with the receiver. The third could sit in a drawer, a satellite, or a crate on another continent, and the pair could still burn a pattern into it if the geometry was close enough. Conventional intercept equipment watched spectrum, cables, and timing. It did not watch an unpowered die in a box. There was no emission to collect in the usual sense, only a scar that appeared on a part that had never been addressed.

The agency did not pretend the channel was free. Each forced write ate lifetime. Em's thermal models, rerun with their failure library, put a matched segment at roughly twenty thousand hours of message traffic before the line was too damaged to serve as a clean bit. After that the chip was scrap, useful only as a record of what had already been sent. They would not stretch it. They would stock replacements.

Warehouses were the real design. Sets of three, and then sets of replacements for those sets, serialized so that a burned die could be swapped without breaking the match. A capital kept a vault. An embassy kept a safe. A ship kept a locker. When one chip aged out, a twin from the same original neighborhood took its place, and the other two ends of the link updated their maps. The plan on the whiteboard used the phrase worldwide first, then added a second line in smaller type: and beyond, if a matched set could be carried.

Em read that line twice. Lithography did not care about distance. Match cared about the wafer. If three dies from the same shot could be kept as a family, the family could be split across any gap a vehicle could cross. A message would still arrive as damage. No one outside the set would see it leave.

She went back to the lab because they still needed her to bin the run. The racks were the same. The red blocks were the same. The difference was the label on the crate at the door, which no longer said scrap. It said hold. Em logged the next neighborhood, three dies and their spares, and wrote the hour count in the corner of the sheet: 20000, then replace. Outside, the full run was already being scheduled. Inside, the chips waited in matched silence, identical enough to burn one another from across a room, and soon from much farther than that.