Microbump Thermal Cycling Fatigue
Microbump thermal cycling fatigue is the slow mechanical failure of the thousands-to-millions of solder microbumps joining a GPU die to its interposer or substrate, driven not by any single overload but by the die, interposer, and substrate expanding and contracting by different amounts every time the chip heats up and cools down. At a high level, the damage accumulation looks like this:
A few things are worth knowing about why this is tracked as its own reliability budget, separate from both electromigration and one-time mechanical overstress:
It is a fatigue problem, not an overload problem. No single thermal cycle comes close to breaking a microbump outright — the solder deforms elastically within its normal operating range every time. The damage is cumulative plastic strain that builds joint by joint, cycle by cycle, which is why it is modeled with the Coffin-Manson relation (cycles-to-failure scales with a negative power of strain amplitude) rather than a simple stress-threshold pass/fail test: a bump that survives a thousand cycles can still fail on its ten-thousandth, and the only way to know where that line sits is accelerated thermal-cycling qualification, not a single stress test.
Distance to the neutral point decides which bumps fail first, every time. Because the die and substrate expand around a shared center point, any given microbump's shear strain scales with how far it sits from that neutral point — corner and edge bumps on a large reticle-size GPU die see far more strain per thermal cycle than bumps near the die center, which is why package designers deliberately route dummy or under-fill-reinforced bumps at the corners, and why the largest dies (the ones most valuable to protect) are also the ones with the worst DNP geometry working against them.
Underfill exists specifically to fight this mechanism, not to protect against dust or moisture. The epoxy underfill dispensed into the gap beneath a flip-chip die after bumping mechanically couples the die and substrate together so they share strain more evenly and constrains each bump from shearing independently — without it, thermal-cycling fatigue life on a large die would be a fraction of what qualification requires. Underfill material choice (its own CTE and modulus) is itself tuned against this same fatigue model, making it as much a reliability component as a protective one.
Read microbump fatigue through a *cumulative-strain-by-position* lens rather than a *single-weak-joint* lens: every thermal cycle does a little damage to every bump, strain scales predictably with distance from the package's neutral point, and the entire packaging toolkit around it — underfill, corner reinforcement, accelerated cycling qualification — exists to push the Coffin-Manson cycles-to-failure number for the worst-positioned bump out past the chip's expected service life.