Advanced packaging pushes warpage metrology beyond a single flatness number

Wooptix, an optical measurement company, says chip packaging needs a better way to check surface flatness. Most packages today pass or fail on a single flatness number, but Wooptix argues that number hides too much. A warped surface can come from several problems at once, not one simple bend.

Adam Chuang, Wooptix's VP of business development for optics, said flatness is really "a composition of multiple different length scales that contribute from the process itself." A surface can bow across the whole package, tilt at individual dies, or turn uneven from steps like bonding, polishing, or heat mismatch, and one number can't tell those apart. His fix: map the whole surface's shape instead of taking a single reading.

Several trends are forcing the change. AI and high-performance chips are getting bigger and moving to new organic and glass materials. Stacking chiplets and memory mixes materials that expand differently when heated. And a newer joining method, hybrid bonding, leaves almost no room for error on flatness. Bigger panels raise the stakes too: ASE already runs an automated line at 310mm by 310mm, beyond the round 300mm wafers used for decades.

Hybrid bonding is the strictest case because there's no soft material at the joint to absorb imperfection. The surfaces are mostly oxide, which must be extremely flat to bond. As Mike Kelly, VP of chiplets and FCBGA integration at Amkor, put it, "oxides want to be bone flat... The copper is just along for the ride." Press too hard during later heating and the copper contacts deform, so a surface that looks fine at one step can still fail the next.

Glass carriers are gaining ground partly because their expansion is tunable. Wiwy Wudjud, a program manager at ASE, noted "glass has a tunable CTE, ranging from 3.4 to 10.6 ppm per degree." But a full stack still mixes materials that expand differently, so Kelly said designers need to spread those materials evenly through the package.

Simulation has its own roadblock. Lang Lin, a principal product manager at Synopsys, said stress models need real material data measured from the actual substrates, but that data reflects suppliers' secret recipes and they're reluctant to share it. Without it, simulations fall back on generic data that shows trends but can't predict actual outcomes.