Defect-Aware Chiplet Matching For Heterogeneous Integration Yield (ASU, Ampere)

Chip Manufacturing

ASU and Ampere Unveil DT-MATCH Chiplet Matching Framework

ASU and Ampere researchers published DT-MATCH, a defect- and topology-aware chiplet matching framework using ILP and GRASP methods to maximize functional SoC yield.

By
Rebecca Stone
Filed
Channel
Chip Manufacturing
Read
3 min read

Arizona State University and Ampere Computing researchers published a chiplet-matching framework on November 14, 2026, that aims to raise functional system-on-chip yields by pairing defective-but-repairable chiplets more intelligently before assembly. The paper, "DT-MATCH: Defect- and Topology-Aware Chiplet Matching for Yield Optimization in Heterogeneous Integration," appeared as arXiv:2611.19544.

The five authors — Xuanyi Tan, Sanmitra Banerjee, Dhruv Thapar, Sreejit Chakravarty, and Krishnendu Chakrabarty — tackle a problem the chiplet era created. Heterogeneous integration and advanced packaging let designers build SoCs from multiple dies. But they also, as the paper states, "introduce new yield challenges arising from manufacturing defects and complex die-to-die (D2D) interconnect constraints."

The core issue is matching. Lane-level defects, limited spares, and strict structural compatibility requirements, in the authors' words, "impede the efficient assembly of functional systems from defective but repairable chiplets." A chiplet with a few bad lanes is not scrap — it is a candidate, if you can find it the right partner die.

What does DT-MATCH actually do?

DT-MATCH maximizes the number of functional SoCs assembled under realistic constraints. It offers two complementary solution paths:

  • An integer linear programming (ILP)-based approach, implemented with a batching strategy that approximates the exact formulation.
  • A greedy randomized adaptive search procedure (GRASP)-based approach, which trades guaranteed optimality for substantially faster runtimes.

Both share a preprocessing pipeline that shrinks the problem before either solver runs. That matters because chiplet matching at production scale is a combinatorial problem — die populations run into the thousands, and naive exact optimization does not survive contact with a fab's throughput requirements.

How do the two methods compare?

The team validated both on industrial designs, not synthetic benchmarks. The results land where you would expect. "As expected, the ILP-based approach achieves superior solution quality," the authors write. The GRASP variant "leads to lower runtimes while still achieving solution quality that is comparable to the ILP-based solution."

That framing is the practical takeaway: fabs and OSATs can choose their point on the quality-versus-speed curve. A high-mix, high-volume packaging line might accept GRASP's near-optimal matches to keep assembly lines moving. A lower-volume, high-margin part might justify the ILP's batching strategy to squeeze out the last points of yield.

The authors sum up the contribution plainly: "Together, these methods provide a trade-off between solution quality and efficiency for practical chiplet assembly optimization."

Why does this matter for the chiplet supply chain?

As AMD, Intel, and Arm-based designers push multi-die architectures into mainstream products, die-level yield recovery has become a commercial lever, not an academic curiosity. Known-good-die economics are already hard; heterogeneous integration makes them harder, because a finished SoC is only as viable as the worst-constrained link between its chiplets. A framework that systematically harvests repairable defective dies instead of discarding them attacks cost directly at the packaging step.

The involvement of Ampere Computing, the server-CPU designer, signals where this research is heading: large die-to-die interconnect fabrics in datacenter silicon, where lane-level spare management across many chiplets is precisely the constraint DT-MATCH models.

The ASU-Ampere team leaves open how the framework performs under full production-scale die populations beyond the reported industrial cases — the natural next test as chiplet-based designs move from premium server parts toward volume platforms.

Original: arxiv.org

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,092 articles

Related articles

« Previous article