TL;DR
As chip designs shrink, predicting timing delays caused by "crosstalk" between adjacent wires becomes crucial but challenging. This paper introduces a two-step machine learning approach that accurately predicts these delays during the routing phase, avoiding the pessimism and long iteration times of traditional methods.
What Problem Does This Paper Address?
In modern integrated circuits, wires are packed so closely together that the electrical signals on one wire can interfere with signals on neighboring wires. This phenomenon, known as crosstalk, can significantly alter the timing of signals, leading to either slowdowns (increasing delay) or speedups. Accurately predicting this crosstalk-induced delay is vital for ensuring a chip functions correctly and meets its performance targets.
Current methods for predicting crosstalk delay often rely on post-routing information and parasitic extraction, which are computationally expensive and only available late in the design cycle. This leads to overly pessimistic timing estimates during the routing phase, forcing designers to add extra timing margins. These margins can result in longer design times due to repeated routing iterations and can even lead to suboptimal chip performance. The core problem this paper tackles is the need for an accurate, efficient, and early-stage prediction of crosstalk-induced delay during the routing process itself.
Methods
The paper proposes a novel two-step machine learning approach for crosstalk-aware timing prediction.
- Crosstalk Filtering (Step 1): This crucial first step aims to identify which physically adjacent nets are actual crosstalk aggressors and which are not. It leverages timing-window-related features, such as the relative arrival time between signals on adjacent nets (Askew), the length of coupling segments (LSI), and the width of coupling segments (WSI). By analyzing the overlap of timing windows, this step filters out nets that are physically close but whose signals don't overlap in time, thus avoiding unnecessary and pessimistic delay calculations. This filtering significantly reduces the complexity for the subsequent step.
- Delay Prediction (Step 2): For the filtered nets, this step predicts the crosstalk-induced delay. It integrates physical features (like interconnect width, thickness, dielectric properties, and length) with timing features (driver cell delay, input/output transition times). The model uses a combination of machine learning algorithms, including Multilayer Perceptrons (MLP), Random Forests (RF), and LightGBM, trained independently for classification and prediction tasks. The goal is to predict the stage delay, which includes both the no-crosstalk delay and the crosstalk-induced delta delay.
This two-step approach is designed to be integrated directly into routing tools, enabling early and accurate timing estimations without requiring post-routing results.
Datasets
The paper uses benchmark designs from OpenCores for training and evaluation. Specifically, 11 designs from the open-source benchmark suite were synthesized, placed, and routed using a foundry 14nm FinFET PDK. The training data was generated from approximately 1.91 million training samples, with 70% used for training and 30% for testing. The ground truth for comparison was obtained using commercial tools like StarRC for parasitic extraction and PrimeTime-SI for static timing analysis.
Key Insights and Ideas
- Timing Window is King: The paper strongly emphasizes that crosstalk delay is not solely determined by physical proximity or coupling capacitance. The timing overlap between signals on adjacent wires is a critical factor. Ignoring this can lead to significant over-pessimism.
- Filtering is Efficient: The two-step approach, particularly the first step of filtering non-critical crosstalk pairs, is a key enabler for efficiency. It drastically reduces the number of nets that need detailed crosstalk delay analysis, making the overall process faster and more manageable.
- Early Prediction Pays Off: By performing this analysis during routing, designers can make more informed decisions earlier in the design flow, potentially avoiding costly reruns and improving overall chip performance.
- Feature Engineering Matters: The successful integration of both physical and timing-window-related features is crucial for the model's accuracy. This highlights the importance of understanding the underlying physics of crosstalk when designing ML models for EDA.
Key Results
- High Accuracy: The proposed method achieves a match rate of over 99% for identifying crosstalk nets compared to commercial tools on OpenCores benchmarks.
- Improved Prediction: The model demonstrates more accurate delay prediction results than other state-of-the-art methods, as evidenced by high R² scores (e.g., up to 0.9983 for LightGBM in the second step).
- Reduced Pessimism: The timing-window-based filtering effectively removes "False SI nets" (FSI), which are nets with coupling capacitance but no timing overlap, thus reducing the pessimism in delay predictions.
- Competitive Runtime: While slightly slower than some traditional methods, the runtime is considered acceptable for integration into routing tools, offering a significant speedup compared to full post-routing analysis.
Limitations and Future Work
- Runtime: The paper acknowledges that their crosstalk-aware method is "a little slower" than some other approaches, though still acceptable. Future work could focus on further optimizing the runtime.
- Integration: While the paper states the model can be integrated into existing routing tools, the practical implementation details and performance gains in a full-flow scenario would be an area for further exploration.
- Model Generalization: The current model is trained on specific benchmarks and technology nodes. Testing and fine-tuning on a wider range of designs and technologies would be beneficial for broader applicability.
- Dynamic Behavior: The current model focuses on static timing analysis. Exploring its application to more dynamic scenarios or considering signal integrity issues beyond simple delay could be a future direction.
Summary updated