Defect-Aware Chiplet Matching Boosts Heterogeneous Integration Yield

Defect-Aware Chiplet Matching for Heterogeneous Integration Yield


ASU and Ampere researchers propose a matching strategy that pairs chiplets by defect profile to raise assembly yield in heterogeneous integration.


Overview


Heterogeneous integration—assembling multiple chiplets into a single package—has become a mainstream path to scaling performance and cost in advanced semiconductor design. By 2026, chiplet-based architectures are standard across data center CPUs, GPUs, AI accelerators, and networking silicon, with die-to-die interconnects such as UCIe 2.0 and 3D stacking now widely deployed. Yet yield remains a persistent challenge: each individual chiplet must be known-good before assembly, and even then, defects that escape test can degrade the final package.


Researchers from Arizona State University (ASU) and Ampere Computing have proposed a defect-aware chiplet matching methodology that pairs chiplets according to their defect profiles to improve overall heterogeneous integration yield.


The Yield Problem in Chiplet Assembly


Traditional known-good-die (KGD) screening tests each die individually. However, in a multi-chiplet package, the interaction between dies matters. A chiplet with marginal defects may still function, but when combined with another marginal die, the assembled system can fail specification. Conversely, pairing a strong die with a weaker one may be perfectly acceptable.


The core insight of defect-aware matching is that chiplets are not interchangeable—their measured defect characteristics influence how well they will perform together in a package.


Matching Strategy


Rather than treating all passing chiplets as equivalent, the ASU/Ampere approach characterizes each chiplet's defect signature and uses that information to guide pairing decisions during assembly. This allows manufacturers to:


  • Maximize assembled yield by avoiding combinations of marginal dies that would push the package out of spec.
  • Recover value from lower-grade chiplets by pairing them with stronger counterparts.
  • Reduce scrap and improve cost efficiency in high-volume heterogeneous integration flows.

2026 Context


As chiplet ecosystems mature, yield optimization has moved from a back-end packaging concern to a first-class design and manufacturing problem. Industry consortia continue to refine KGD standards, and advanced packaging capacity—2.5D interposers and 3D hybrid bonding—has expanded significantly. Defect-aware matching complements these efforts by adding a data-driven decision layer at the assembly stage.


With UCIe-based multi-die designs now common and AI accelerator demand driving unprecedented packaging volumes, techniques that squeeze additional yield from existing die inventory offer direct economic benefit.


Implications


Defect-aware chiplet matching represents a shift toward treating assembly as an optimization problem rather than a simple pick-and-place operation. For design teams and OSATs, it suggests closer integration between test data, defect characterization, and packaging decisions—an approach likely to become more valuable as chiplet counts per package continue to rise.

via Semiconductor Engineering

Related