Defect-Aware Chiplet Matching For Heterogeneous Integration Yield (ASU, Ampere)
Original reporting by Semiconductor Engineering

DT-MATCH is a novel computational framework designed to optimize the assembly of complex chiplet-based systems by intelligently matching defective components to maximize overall manufacturing yield. As the semiconductor industry moves towards heterogeneous integration, where multiple specialized "chiplets" are combined into a single system-on-chip (SoC), new challenges arise. Manufacturing defects are inevitable, but if repairable, these defective chiplets can still be used, provided they are matched correctly according to stringent die-to-die interconnect requirements and limited repair resources. The difficulty lies in efficiently finding the best combinations to produce the maximum number of functional SoCs from a pool of imperfect chiplets.
Optimizing Chiplet Assembly
Researchers from Arizona State University and Ampere Computing developed DT-MATCH to address this critical yield bottleneck. Their framework offers two complementary approaches: an Integer Linear Programming (ILP) method, which prioritizes optimal solution quality, and a Greedy Randomized Adaptive Search Procedure (GRASP) method, which offers significantly faster runtimes while still delivering comparable, albeit not guaranteed optimal, results. Both methods employ preprocessing to streamline the problem, providing a flexible trade-off between the pursuit of absolute perfection and the demands of practical, high-volume manufacturing. This innovative work, detailed in their November 2026 paper, offers a crucial tool for enhancing the economic viability of next-generation chip designs.
The DT-MATCH framework marks a pivotal advancement in optimizing the assembly of chiplet-based systems, a cornerstone of modern high-performance computing. By providing both precise ILP-based and efficient GRASP-based solutions, researchers at Arizona State University and Ampere Computing have delivered critical tools to navigate the intricate challenges of heterogeneous integration. This capability to effectively manage manufacturing defects and strict die-to-die interconnect requirements directly addresses a principal hurdle in scaling advanced packaging techniques. This research brings the industry closer to fully realizing the transformative potential of chiplet designs, ensuring that highly specialized and powerful processors can be reliably produced even from individual components exhibiting minor imperfections.
Paving the Way for Advanced Computing
The implications of DT-MATCH extend significantly beyond immediate yield improvements. This work lays foundational groundwork for the continued proliferation of chiplet architectures, which are increasingly essential for developing next-generation AI accelerators, high-performance computing platforms, and specialized hardware tailored for diverse workloads. By making the assembly of complex, multi-chiplet systems more efficient and less susceptible to costly failures, DT-MATCH significantly lowers barriers to innovation. It empowers designers to explore more intricate heterogeneous designs, pushing the boundaries of performance, power efficiency, and functionality. As the global demand for custom silicon continues to escalate across industries, solutions like DT-MATCH will prove indispensable, ensuring that the promise of modular, high-yield chip manufacturing translates into a consistent supply of powerful, reliable AI hardware and advanced computing systems critical to our technological future.