Intel saved $500M by routing multidie chip assembly instead of matching dies

Semiconductor design shifted from large monolithic dies to multidie packages, stacking several smaller components to integrate more functionality. That tripled the number of parts to manage and turned assembly decisions into an interdependent, nonlinear problem: the quality of any assembly depends on which dies are grouped with which.

Intel named this the sort-assemble-blend routing problem (SABR-P) and solved it with a stack of genetic algorithms, Monte Carlo simulation, linear programming and machine learning. The framework sorts incoming materials into compatibility categories and then routes each die into an assembly so the whole portfolio's output is optimized rather than any single pairing.

Runs in practice have saved Intel more than $500 million so far, and the approach is now the standard for analyzing and optimizing Intel's multidie products.

Treat sort, assemble and blend as one routing problem

Rival approaches optimize each die or each pairing; SABR-P treats the full pipeline as one routing decision, so trade-offs between dies surface instead of being hidden.

The hybrid solver lets nonlinear manufacturing realities (yield, compatibility, blending) be represented directly, which pure linear or heuristic methods cannot capture.

Once the framework proved itself across product lines, it became the default tool rather than a one-off study, compounding the savings.

When a product turns into a sum of interchangeable parts, stop matching parts pairwise — model the whole flow as a routing problem and let the optimizer choose who goes with whom.

The SABR-P framework won the 2023 Daniel H. Wagner Prize for Excellence in the Practice of Advanced Analytics and Operations Research. Intel has adopted it as the new standard for multidie product analysis and optimization for the foreseeable future, with savings exceeding $500 million documented in the INFORMS Journal on Applied Analytics.