Sleep transistor distribution in row-based MTCMOS designs

Chanseok Hwang, Peng Rong, Massoud PedramACM Great Lakes Symposium on VLSI 2007 · DOI

TL;DR

This paper tackles the challenge of minimizing performance degradation in MTCMOS circuits caused by the placement of sleep transistors. They introduce an optimal algorithm that intelligently distributes these transistors across standard cell rows, leading to significant reductions in critical path delay compared to existing methods.

What Problem Does This Paper Address?

Modern integrated circuits face a growing problem with leakage power, which can rival dynamic power consumption. Multi-Threshold CMOS (MTCMOS) is a popular technique to combat this by using high-threshold voltage (high-Vth) "sleep transistors" to cut off power to logic blocks during standby mode. However, during active mode, these sleep transistors introduce a small resistance into the virtual ground network. This resistance, combined with the interconnect resistance of the virtual ground network itself, can cause voltage drops. These voltage drops reduce the effective gate-overdrive voltage of logic cells, leading to increased propagation delays and overall performance degradation. Existing methods for placing these sleep transistors often place them at the corners of cell rows or distribute them uniformly without considering the impact of their placement on the virtual ground network's resistance. This paper aims to optimize the placement of these sleep transistors to minimize this performance penalty.

Methods

The authors propose an optimal algorithm for distributing sleep transistors (STs) within standard cell rows. The core of their approach is to formulate the problem as minimizing the maximum performance loss across all logic cells in a row. This performance loss is modeled as a function of the distance between a logic cell and its nearest sleep transistor, considering the virtual ground interconnect resistance.

They define a "performance loss" metric for each cell and then aim to minimize the maximum performance loss across all cells in a row, subject to constraints on the current each sleep transistor can handle. Their approach leverages dynamic programming to efficiently search for the optimal placement of sleep transistor cells (STCs) on each row. The algorithm takes as input the coordinates of logic cells and the number of STCs to be placed, and outputs the optimal locations for these STCs and the minimized performance degradation.

Datasets

Not discussed in the paper. The authors mention using ISCAS benchmark circuits and generating optimized gate-level netlists using SIS and a timing-driven placer.

Key Insights and Ideas

  • Placement Matters: The physical placement of sleep transistors is not just about area but critically impacts performance due to the virtual ground network's resistance.
  • Distance is Key: The performance degradation of a logic cell is primarily influenced by its proximity to the nearest sleep transistor.
  • Dynamic Programming for Optimization: A dynamic programming approach can efficiently find the optimal distribution of sleep transistors, avoiding the impracticality of brute-force search for larger circuits.
  • Balancing Current: It's crucial to balance the current flowing through each sleep transistor to prevent localized voltage drops that disproportionately affect nearby cells.

Key Results

The proposed Optimal ST distribution technique (OSTD) was compared against two other methods: Corner-based ST distribution (CSTD) and Uniform ST distribution (USTD).

  • Critical Path Delay Reduction: OSTD reduced the critical path delay by an average of 11% compared to CSTD.
  • Virtual Ground Bounce: OSTD reduced virtual ground bounce by approximately 50% (from 250mV to 180mV) in simulations of the virtual ground line.
  • Comparison with USTD: OSTD achieved an average critical path delay reduction of 3.84% compared to USTD.
  • Wire Length: OSTD resulted in a small average increase in total wire length (0.7%) due to layout adjustments needed to resolve overlaps between sleep transistors and logic cells.

Limitations and Future Work

  • Limited Benchmarks: The authors note that they were limited to small circuit benchmarks due to HSpice simulation capacity. They expect the advantages of OSTD to become more pronounced with larger circuits.
  • Future Work: The paper implicitly suggests that extending this optimal placement strategy to more complex circuit layouts and considering other performance metrics beyond critical path delay would be valuable future research directions.

Summary updated

BibTeX
@inproceedings{hwang2007sleep,
  author = {Chanseok Hwang and Peng Rong and Massoud Pedram},
  title = {{Sleep transistor distribution in row-based MTCMOS designs}},
  booktitle = {ACM Great Lakes Symposium on VLSI},
  year = {2007},
  doi = {10.1145/1228784.1228786},
}
Figure