TL;DR
As transistors shrink, Line-Edge Roughness (LER) becomes a major source of performance variation. This paper introduces a novel characterization methodology that accounts for LER by simulating non-rectangular gate shapes, enabling more accurate prediction of device performance and suggesting LER tolerance limits for sub-45nm standard cells.
What Problem Does This Paper Address?
As semiconductor devices continue to shrink, particularly below the 45nm node, the impact of manufacturing variations on circuit performance and yield becomes increasingly critical. One such variation, Line-Edge Roughness (LER), refers to the jaggedness of the edges of transistor gates. Unlike other variations that tend to scale down with device size, LER has not decreased proportionally and now represents a significant fraction of the gate length in advanced nodes. This roughness leads to unpredictable fluctuations in device parameters, such as drive current and leakage, ultimately impacting circuit reliability and performance. Existing characterization methods often struggle to accurately model and account for this random LER, especially when combined with systematic lithography variations.
Methods
The authors propose a new cell characterization methodology that integrates random LER into the design flow. The core of their approach involves:
- LER Modeling: They develop an analytical model for LER that can handle arbitrary RMS amplitudes and frequencies. This model is used to generate non-rectangular gate shapes that mimic the effects of LER.
- Print-Image Simulation: Lithography and etch simulations are employed to generate realistic "print-images" of the gate structures, incorporating both systematic lithography variations and the newly modeled random LER.
- Effective Gate Length Extraction: A segmentation-based approach is used to extract the effective gate length and width from these non-rectangular print-images. This method divides the gate into small segments and accounts for diffusion rounding effects.
- TCAD Validation: The accuracy of their LER model and characterization flow is validated using physics-based TCAD simulations. They specifically use a strained silicon process for NMOS transistors in a 45nm standard cell.
- Statistical and Deterministic Analysis: The methodology is designed to be applicable to both statistical and deterministic analysis flows, allowing for a comprehensive understanding of LER's impact.
Datasets
Not discussed in the paper. The authors mention using commercial simulation tools (HSPICE, Sentaurus TCAD) and experimental SEM images for LER frequency data.
Key Insights and Ideas
- LER as a Dominant Factor: For channel lengths below approximately 30nm, LER emerges as a more significant source of fluctuation than random dopants, highlighting its increasing importance in advanced nodes.
- Non-Rectangular Gate Shapes are Key: The paper emphasizes that LER fundamentally alters gate shapes from ideal rectangles to irregular contours. Accurately capturing these non-rectangular shapes is crucial for predicting performance.
- LER's Impact on Current: The driving current of transistors is shown to be highly sensitive to LER, especially as gate lengths become thinner. This sensitivity directly translates to performance variations.
- Systematic vs. Random LER: The methodology aims to disentangle and account for both systematic lithography variations (due to pattern proximity) and random LER, providing a more complete picture of process variation.
- LER Tolerance is Necessary: The work suggests that defining maximum LER tolerance is essential to minimize performance degradation and ensure circuit yield.
Key Results
- LER Amplitude: Experimental results from 45nm process simulations show that the RMS value of LER can be as high as 10% of the nominal line edge.
- Current Variation: In a 2-input NAND cell, the saturation current can vary by as much as 10% due to LER.
- Model Accuracy: The proposed LER model shows excellent agreement with TCAD simulations, with a maximum error of 5.3% and an average error of 1.2% in current variation.
- Delay Sensitivity: Table 1 demonstrates that delay sensitivity to LER varies significantly across different process corners (best, typical, worst), underscoring the need for corner-aware characterization. The sensitivity is generally higher in typical and best-case corners.
- Delay Impact: Figure 9 shows that delay variation is minimal at low LER levels (e.g., <3% CD) but becomes steep as LER increases, due to the exponential relationship between gate length and saturation current.
Limitations and Future Work
- Leakage Current and Stress Effects: The authors mention that their future work will focus on investigating the impact of LER on leakage current and other stress effects.
- Model Complexity: While the analytical LER model is flexible, its integration into complex lithography simulation flows might require further optimization.
- Experimental Data: While TCAD validation is strong, direct experimental validation of the full characterization flow on fabricated devices would further strengthen the findings.
Summary updated