As CMOS technology scales into advanced nodes, modern low-power, high-performance system-on-chip (SoC) designs increasingly rely on both multi-threshold voltage technology and mixed-cell-height standard cell libraries. In such designs, small perturbat...
As CMOS technology scales into advanced nodes, modern low-power, high-performance system-on-chip (SoC) designs increasingly rely on both multi-threshold voltage technology and mixed-cell-height standard cell libraries. In such designs, small perturbations in device threshold voltages (VT) can cause large fluctuations in leakage power, while the increased complexity of legalization for mixed cell heights can cause large deviations from the timing- and routability-driven global placement solution. In this context, this dissertation studies two related optimization problems that arise at different stages of the design implementation flow: (i) recovering leakage power by tuning VT values in timing-closed multi-threshold voltage designs at the post-route stage, and (ii) recovering wirelength and placement quality by refining legalization for mixed-cell-height layouts during the placement stage of the physical design flow.
For the first problem, given a timing-closed design typically at the post-route stage whose leakage power has already been optimized for nominal VT values, we reinterpret VT tuning as a design-side upper-bound and guidance problem. Specifically, we (i) quickly estimate the maximum leakage power reduction achievable by applying small threshold-voltage shifts to each VT type (e.g., RVT/LVT/SLVT for nMOS/pMOS) without violating timing constraints, and (ii) use the resulting continuous optimum to guide practical choices from a discrete pool of VT options provided by the process/library. To this end, we develop a near-optimal approximation framework based on a piecewise-linear programming formulation that embeds delay and leakage sensitivities into linear constraints and objective. By replacing repeated full-chip static timing analysis (STA) over sampled VT configurations with a single linear-programming (LP) solve over piecewise-linear delay and power models, the proposed method avoids slow STA loops while providing fast upper-bound estimation and actionable guidance; it achieves five orders of magnitude speedup over conventional VT tuning flows, while obtaining, on average, 5.45% additional leakage power reduction under the given timing budgets.
For the second problem, we propose a window-based legalization refinement framework tailored for mixed-cell-height designs. Starting from a legal placement delivered by a commercial tool or an MGL-style legalizer, our method iteratively scans windows, generates slots by aggregating candidate cells and adjacent whitespace, and solves a min-cost perfect matching problem between candidate cells and slots. This formulation enables cooperative multi-cell rearrangements and explicitly supports substitution among functionally equivalent cells with different heights, reflecting modern library trends. We further introduce displacement-like priority metrics to guide candidate selection and exploit independent window sets for multiprocessing on multi-core platforms. Experiments on OpenCores benchmark circuits implemented using the ASAP7 7.5-track standard cell library show that, over commercial legalizations, the proposed refinement reduces average displacement, maximum displacement, and half-perimeter wirelength (HPWL) by an average of 25.7%, 52.5%, and 6.6%, respectively, leading to a 38.7% improvement in a composite score inspired by the ICCAD-2017 multi-deck legalization contest, with up to 66.47% score improvement on extremely dense designs.
Taken together, these two contributions provide a unified methodology for improving both power and interconnect quality in advanced-node designs. VT tuning globally reshapes the leakage–delay trade-off at the device and cell level in timing-closed post-route designs, while mixed-height legalization refinement locally rearranges cells and whitespace during placement to recover wirelength and placement metrics and better preserve the intent of the timing-driven global placement.