The 68 mV Reference Shift: Field Notes on a Cracked Power IC Die

Archi4 min read

semiconductorspowerfailuretesting

Contents
  1. Bench Tests Did Not Show It
  2. The Usual Suspects
  3. The Load Transient Told the Story
  4. Thermal Cycling and a Probe
  5. Decapsulation Found the Crack
  6. Why the Die Cracks
  7. The Fix and the Verification

Bench Tests Did Not Show It#

Eight boards from one production batch came back with the same complaint. After about 30 minutes of powered operation, the 3.3V rail dropped to 2.85V. On the bench, the fault did not appear at room temperature. It appeared only after the board was heated with a warm air gun set to 60°C. Cooling the board returned the rail to 3.30V. This was not a random glitch. It repeated for all eight boards.

Field note: always ask what the customer was doing when the fault appeared. In this case, all eight boards were in enclosures with poor airflow. The case temperature reached 78°C. The board power supply was the only common part.

The Usual Suspects#

First I checked the input. The 5V supply stayed at 5.02V with less than 30mV ripple. The inductor was a 10µH shielded part, measured at 9.6µH with no saturation at 1.5A. The output capacitor was a 47µF ceramic with 10µF of effective capacitance at 3.3V, which is normal. No cracked solder joints, no lifted pads, no dark spots under a microscope.

Then I replaced the feedback divider with a pair of 0.1% precision resistors. The fault still appeared. That removed the external divider from the list.

The Load Transient Told the Story#

With a 0A to 1A load step, the output dropped 20mV and recovered in about 40µs. That was normal. The problem was the absolute voltage level, not the transient response. The output was 2.85V regardless of load from 100mA to 1.2A. The regulation loop was working. It was regulating to the wrong target.

A buck converter regulates its output to the reference voltage multiplied by the feedback divider ratio. The reference for this controller was 0.600V. With a divider ratio of 63.4k over 10k, the expected output was 3.30V. For the output to sit at 2.85V, the reference needed to be 0.532V. That is a shift of 68mV.

Thermal Cycling and a Probe#

I put the board in a thermal chamber at 25°C and monitored the output. At 25°C, the output was 3.29V. At 60°C, it dropped to 2.85V. At 40°C, it was back to 3.28V. The transition happened between 57°C and 62°C. It was not gradual. It stepped by about 60mV, which is a switch, not a drift.

Actually, the output did not step instantly. It moved over about 3 seconds. That suggested a crack opening or a connection breaking as the package expanded. I reflowed the part with fresh solder. The fault did not go away. The IC itself was the problem.

Decapsulation Found the Crack#

We sent a failed IC to a failure analysis lab. The lab removed the mold compound from the die. The image showed a crack across a thin-film resistor near one corner of the die. The crack ran from the die edge through a serpentine resistor and stopped at a metal trace. It was about 200µm long.

The resistor was part of the bandgap reference network. The bandgap reference uses a pair of resistors with a specific ratio to set the voltage. When the crack increased the resistance of one leg, the ratio changed, and the output voltage moved. This explained the 68mV shift and the temperature threshold.

Why the Die Cracks#

The package was a DFN-8 with an exposed pad. The PCB had a mounting screw hole 8mm from the IC. The assembly spec for the screw was 3 in-lb of torque. The line used a driver set to 5 in-lb. The extra torque flexed the board enough to put the package under stress. The stress concentrated at the corner of the die, and thermal cycling did the rest.

We measured the board flex with a dial gauge. At 3 in-lb, the board deflected 0.2mm under the IC. At 5 in-lb, it deflected 0.5mm. The package transferred that deflection to the die through the exposed pad. The die is brittle. It cracked.

The Fix and the Verification#

We changed three things. The screw torque was reset to 3 in-lb. A washer was added to spread the load. The PCB layout now has a small slot between the screw hole and the IC, so board flex has a place to stop. We also placed a small damping loop in the feedback network to filter any noise from the input, though that was not the root cause.

We ran 40 boards through 500 thermal cycles from 0°C to 85°C. No failures. We also tested 10 boards with screw torque at 5 in-lb and got one failure after 300 cycles. That matched the original failure rate.

Field note: a power IC can pass every functional test and still fail in the field. The failing mode was not electrical, it was mechanical. The bench test needs to include the mechanical environment, especially when the board has a nearby fastener.

More articles