Abstract
This paper presents a novel stochastic optimization framework for VLSI decoupling capacitance (decap) budgeting. By leveraging Independent Component Analysis (ICA) to model logic-induced and temporal correlations alongside process variations, the approach avoids the over-pessimism of traditional "worst-case" assumptions. Results on industrial benchmarks demonstrate up to a 17x reduction in predicted noise and a 13x optimization speedup compared to standard industrial methods.
Introduction
In high-performance integrated circuits, Power/Ground (P/G) networks are tasked with delivering massive currents while maintaining voltage stability. Significant voltage drops (IR-drop) or inductive noise ($Ldi/dt$) can lead to timing violations or logic failures. To mitigate this noise, designers allocate decoupling capacitors (decaps) across the chip.
Traditionally, decap budgeting has relied on "worst-case" scenarios, where designers assume every port on the chip hits its maximum current peak simultaneously. This assumption is physically unrealistic and leads to significant "over-designing," consuming precious silicon area with unnecessary decaps. This research addresses the "pessimism gap" by proposing a stochastic optimization routine that accounts for the complex correlations inherent in real-world circuit activity.
The Challenge of Correlation
To create a more realistic model, the authors identify three primary sources of correlation that traditional models ignore:
- Logic-induced correlation: The structural connectivity of logic gates means that certain parts of a circuit cannot be active simultaneously.
- Temporal correlation: Activity in one clock cycle is statistically related to activity in subsequent cycles.
- Process variation: Manufacturing inconsistencies, such as variations in effective channel length ($L_{eff}$), alter current behavior across different chips.
By modeling these factors, the optimization framework can distinguish between physically possible current peaks and the statistical impossibility of a global simultaneous peak.
Methodology
The proposed framework utilizes a three-pronged technical approach to solve the variation-aware decap budgeting problem:
Stochastic Current Modeling via ICA
While Principal Component Analysis (PCA) is a common tool for dimensionality reduction, it assumes data follows a Gaussian (normal) distribution. Circuit currents, however, are typically non-Gaussian and "spiky." The authors employ Independent Component Analysis (ICA) to extract independent stochastic variables from correlated, non-Gaussian current signatures, providing a much more accurate representation of digital switching.
Iterative Alternative Programming (IAP)
The core of the framework is a "min-max" optimization problem: minimize the noise for the worst-case input. Because solving this directly is computationally prohibitive, the authors use an Iterative Alternative Programming (IAP) approach that splits the problem into two sub-problems:
- P2: Identifying the worst-case current input for a fixed decap budget.
- P3: Finding the optimal decap budget for that specific "worst" input.
Sequential Programming (sLP/sQP)
To solve these sub-problems efficiently, the framework employs Sequential Linear Programming (sLP) and Sequential Quadratic Programming (sQP). The sQP method, in particular, utilizes second-order sensitivities (Hessians) derived implicitly through Modified Nodal Analysis (MNA) equations. This allows the algorithm to navigate the design space faster than traditional methods that require full circuit simulations at every step.
Experimental Results
The framework was evaluated using four industrial P/G network designs, ranging from 128 nodes (41 ports) to 166,380 nodes (42,146 ports).
Noise Reduction and Accuracy
Compared to the baseline "maximum peak" model, the proposed stochastic model reduced predicted noise by up to 17x. This suggests that traditional methods are vastly overestimating the noise environment. When accounting for $L_{eff}$ process variations, the method reduced mean noise by 4x and $3\sigma$ (worst-case) noise by 13x compared to deterministic models.
Computational Efficiency
The framework achieved a 13x speedup over existing industrial techniques. While the stochastic model is approximately 2.3x slower than the simplest (but highly inaccurate) peak-current model, this is a favorable trade-off given the massive gains in design accuracy and area efficiency. The sQP formulation generally outperformed sLP on larger circuits, demonstrating the value of higher-order sensitivity information in complex design spaces.
Conclusion and Future Work
This research demonstrates that accounting for logic, temporal, and process correlations can reclaim significant silicon area by closing the pessimism gap in decap budgeting. While the current Quadratic Programming formulation is non-convex—potentially leading to local minima—experimental results indicate the solutions are of high quality for industrial applications. Future work may expand this scope to include other operational variations, such as supply noise fluctuations over longer time scales or crosstalk differences across varying input vectors.
Summary updated