Dual-Sourcing a 40nm Auto Chip: The Two-Fab Decision
Archi4 min read
Contents
A chipmaker's choice: pay twice for the same part, or bet on one fab#
Automotive buyers do not purchase microcontrollers the way phone makers purchase processors. They sign supply agreements that run the length of a vehicle model cycle, often seven years, and they write dual-source clauses into those agreements years before anyone needs the second source. For one mid-sized fabless supplier, that clause turned into a two-year internal argument about a 40 nm part that sold for $3.10 and carried a third of company revenue.
The part that paid the bills#
The chip itself was unglamorous: a body-control microcontroller with 512 KB of flash, built at a foundry in Taiwan on a mature node that most of the industry had stopped celebrating around 2018. Volume the previous year was roughly 40 million units. Gross margin sat near 48 percent because the design was old, the masks were paid off, and the test program had been squeezed for a decade.
Then allocation letters replaced purchase orders. The company's largest customer, a tier-one supplier of door and seat modules, sent a formal notice: without a documented second source qualified before the next model year, the business would move to a competitor that could show two fabs on its paperwork.
Two fabs, one mask set#
Porting a 40 nm design between foundries is not a copy and paste job. A different process flavor means different design rules, different standard cells, and in this case a different SRAM compiler, which forced the memory blocks to be regenerated. A fresh mask set at that node runs about $1.5 million. The design team estimated nine engineer-months to re-time the critical paths and re-run timing closure at the corners the automotive spec demands.
Qualification was the longer pole. AEC-Q100 Grade 1 requires 1,000 hours of high-temperature operating life at 125C, plus temperature cycling, biased humidity, and electromagnetic testing. Industry cycles for that package typically run 12 to 18 months, and the clock only starts once silicon comes back from the second fab.
What finance saw#
Masks at $1.5 million. Engineering near $900,000. Another $300,000 to port the test program onto a second probe card set. Qualification hardware, bench time, and the production part approval file pushed the total to roughly $4 million before a single shippable unit existed.
Against that: a contract worth about $40 million a year, and a second foundry quote that came in 6 percent under the incumbent on wafer price. The counterargument came from operations. Splitting 40 million units across two fabs destroys the volume tier at the primary supplier, doubles the wafer starts, and doubles the reliability monitoring overhead. Two dice are never identical, and the analog blocks differed enough that the customer's calibration engineers would need extra margin on the ADC spec, which meant re-validating every module that used the part.
The middle path nobody liked at first#
A third option emerged from the operations team: qualify the second source, but keep the volume split near 85/15 rather than moving whole product lines. That keeps the primary tier intact while satisfying the paperwork. The catch is a die bank. Second-fab wafers have to be built periodically anyway, because a line that sits idle for two years loses its qualified status and the requalification clock restarts.
The company budgeted several hundred wafers a year for that purpose, roughly $2 million of inventory aging in a bonded warehouse in Penang, rotated on an 18-month cycle. Finance initially treated the write-offs as pure waste. Sales treated them as the cost of the contract.
The 14 months that followed#
They qualified. It took 14 months, two months longer than the optimistic plan, and cost about 11 percent over budget after one lot failed temperature cycling and had to be rebuilt with a revised passivation step.
The contract renewed. Volume did not grow. For a while the second fab supplied about 6 million units a year, enough to stay alive and not enough to feel efficient.
Then the primary fab lost a gas supply for six weeks. Allocation notices went out across the industry. Because the die bank existed and the second source was already approved in the customer's system, the door module line kept running with a two-week gap instead of a five-week stop. The finance analyst who had argued hardest against the inventory wrote the internal case study.
What changed in the design process#
Any part above $10 million in annual revenue on a multi-year contract now gets a dual-source plan at architecture review, not after an allocation letter arrives. The analog team builds in wider trim ranges so a second foundry has somewhere to land. Procurement keeps a rolling quote from two fabs even when there is no intention to switch.
The company still writes off around $180,000 a year in expired second-source inventory. Nobody has proposed canceling that line item since the gas outage.
More articles