Photolithography Oxide Masking 1957 Wafer Level Parametric Test
# Wafer-Level Parametric Test: Measuring Hundreds of Devices Before Any of Them Has Been Packaged
## 1. Why It Is Cheaper to Find a Bad Device Than to Package One
This step probes every die on the finished wafer while it is still one piece of silicon — contacting each device's pads, measuring its parameters, and marking it pass or fail, conventionally with an ink dot on the failures — and it exists purely for economic reasons that follow directly from Step 15's arithmetic: packaging is expensive and performed one device at a time, so discovering a bad die now costs far less than discovering the same problem after it has already been packaged. The choice between testing before or after packaging comes down to a simple comparison of two costs, and once a wafer carries hundreds of die rather than one, that comparison stops being close:
Probing first wins whenever $C_{\text{probe}}$ is less than $(1-Y)\,C_{\text{package}}$ — that is, whenever the cost of probing is less than the packaging that would otherwise be wasted on devices that were never going to pass. In this era, with modest yields and largely manual packaging, that inequality holds comfortably, which is the entire reason wafer probe became standard practice rather than an occasional convenience.
## 2. Real Diagram: A Map of Failures Is a Diagnosis, Not Just a Tally
Marking pass or fail on every die produces something more useful than a yield number — it produces a spatial map, and the shape of that map carries information about which earlier step is responsible. A ring of failures at the wafer edge implicates the oxide uniformity problem from Step 2; a contiguous band implicates the registration error from Step 5 or Step 15; scattered singles implicate the random defect density from Step 9.
## 3. Why a Spatial Map Is a Fundamentally New Kind of Measurement
The 1956 process also ran an electrical test — Step 19 of that series measured cutoff frequency and current gain to confirm the drift field actually worked. But that test examined one finished, packaged device at a time, and its output was a single number per device, with no relationship to any other device on the wafer it came from. This step tests hundreds of unpackaged devices in their original positions and produces a spatial map, which carries information no per-device test could ever generate: the correlation between a failure and its physical location. Batch processing did not just lower the cost of testing a device; it created a diagnostic signal — spatial correlation across many devices at once — that a process making one device at a time could not produce even in principle, no matter how carefully each device was measured.
Step 16 does not just separate good die from bad; it reads back, in the shape of the failures alone, which of the preceding fifteen steps is the one actually responsible.