Early-Stage Planning of Switched-Capacitor Converters in a Heterogeneous Chip

Leilei Wang, Lu Wang, Cheng Zhuo, Pingqiang ZhouIEEE Access 2020 · DOI

Introduction

Modern heterogeneous chips, such as the Apple A12 or Kirin 980, integrate a diverse array of cores—including CPUs, GPUs, and neural engines—each with distinct voltage and current requirements. To maximize energy efficiency, these chips increasingly rely on on-chip voltage regulators. Switched-Capacitor Converters (SCCs) have emerged as a popular choice due to their high efficiency and relative ease of integration compared to inductive regulators.

However, SCCs are not without overhead. Integrating them involves significant costs in terms of control circuitry, routing resources, and physical area. Consequently, it is often impractical to provide every core with its own dedicated SCC. This necessitates a complex optimization problem: how to group cores to share regulators while minimizing power loss and how to distribute limited on-chip capacitance effectively.

System-level model of a 4-core chip with integrated SCCs and MIM capacitance.

The Challenge of Heterogeneous Power Delivery

The primary difficulty in early-stage planning for SCCs lies in the trade-off between granularity and overhead. While more regulators allow for finer power management and less "over-supply" waste (where a core receives a higher voltage than it needs because it shares a regulator with a higher-demand core), the fixed integration penalty eventually outweighs the efficiency gains. Furthermore, on-chip "flying" capacitance—typically implemented using Metal-Insulator-Metal (MIM) capacitors—is a finite global resource that must be distributed across all active converters to minimize total system loss.

Proposed Planning Framework

The researchers introduced a two-step framework designed to automate and optimize this planning process:

1. Core Grouping

The framework employs a greedy hierarchical clustering approach to determine which cores should share a regulator. The chip is represented as an adjacency graph where nodes are cores and edge weights represent the "extra power loss" incurred if those cores are merged. The algorithm iteratively merges adjacent cores with the smallest weight until the desired number of SCC groups is reached.

Illustration of core grouping where four cores are clustered into two SCC groups.

2. SCC Optimization

Once cores are grouped, the framework optimizes the SCC parameters using a Mixed-Integer Nonlinear Programming (MINLP) approach. This step decides:

  • Capacitance Allocation: How much of the total available MIM capacitance each SCC receives.
  • Ratio Selection: Which conversion ratio (e.g., 2:1, 3:2, 4:3) each SCC should utilize.

To maintain computational efficiency, the authors simplify the MINLP by enumerating viable ratios for each group and solving the resulting convex sub-problems using a CVX solver. A penalty term is included to account for the overhead of each SCC, allowing the framework to identify the optimal number of converters for a specific chip layout.

Key Insights and Results

The study yielded several critical insights into on-chip power delivery:

  • Capacitance as a Zero-Sum Game: Traditional methods often allocate capacitance based on area or current. This research demonstrates that holistic optimization—shifting capacitance to the SCCs that "need it most" to minimize total system loss—yields significantly higher efficiency.
  • Efficiency Gains: In a 4-core benchmark, the optimized planning improved energy efficiency from 75.3% to 81.9% compared to standard allocation methods. Total power loss for the converters was reduced from 243.92 mW to just 106.84 mW.
  • Optimal Scaling: The framework successfully identified the "sweet spot" for regulator count. For a 16-core benchmark, the optimal number of SCCs was found to be three, balancing management granularity against integration penalties.

Conclusion

By intelligently grouping cores and optimizing the allocation of limited on-chip resources, this framework provides a robust methodology for early-stage chip design. It moves beyond simple area-based heuristics to a system-wide optimization that significantly enhances the energy efficiency of modern heterogeneous processors.

Summary updated

BibTeX
@article{wang2020early,
  author = {Leilei Wang and Lu Wang and Cheng Zhuo and Pingqiang Zhou},
  title = {{Early-Stage Planning of Switched-Capacitor Converters in a Heterogeneous Chip}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2986335},
}
Figure