Worst-case transient excitation determination for heterogeneous-integration power delivery

Determine the worst-case time-varying current patterns that produce the largest voltage drop or worst thermal profile across package-to-chip power-delivery-network interfaces in 3D heterogeneous-integration systems, particularly for highly dynamic AI workloads.

Background

The paper identifies worst-case excitation analysis as a longstanding challenge in power-delivery design. The difficulty arises because different current patterns can produce worst-case behavior in different regions, and because realistic worst cases are generally time-varying rather than static, requiring simultaneous consideration of spatial and temporal behavior.

This problem is especially important for 3D heterogeneous-integration systems supporting AI workloads with microsecond- or submicrosecond-level current transients. Existing vectorless approaches can be pessimistic and do not capture time variation, whereas vectored approaches depend on available workload traces and may fail to find the genuine worst case.

References

An open problem over many decades has been that of determining the worst-case current patterns that result in the largest voltage drop, or the worst thermal profile.

— Toward Multi-kW Power Delivery Methodologies for Advanced 3D Heterogeneous Integration  (2609.24904 - Yue et al., 21 Sep 2026) in Section IV-C, “Determining worst-case excitations”

Scaling such methods to high power levels and fast response times remains an open problem.

— Toward Multi-kW Power Delivery Methodologies for Advanced 3D Heterogeneous Integration  (2609.24904 - Yue et al., 21 Sep 2026) in Section 5.2, “Efficiency optimization and IBV selection”

Several open problems remain, e.g., (a)~regulator placement under transient loads; (b)~simultaneous optimization of the PDN topology, decap insertion, and regulator placement; (c)~regulator placement that is conscious of the regulation capabilities on individual chiplets; (d)~thermally aware placement that keeps regulators away from the hot spots; (e)~ML-assisted strategies for solving all of these problems.

— Toward Multi-kW Power Delivery Methodologies for Advanced 3D Heterogeneous Integration  (2609.24904 - Yue et al., 21 Sep 2026) in Section 5.4, “Regulator placement”