IC Verification Using ML Waveform Prediction Across Timescales
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern integrated circuits, both analog and digital, are complex and time-consuming to design and test due to the difficulty in accurately synchronizing detailed models across various timescales, especially for analog waveforms, which are continuous and specified as limits rather than binary values, making equivalence checking and characterization challenging.
Innovation Solution
A method involving machine learning models is employed to predict output signal values of integrated circuits by training on simulated signal values, comparing them with actual values for validation, using techniques like linear and non-linear regression, and cascading models to enhance prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If detailed modeling is used across all timescales to ensure accuracy, then measurement precision is improved, but device complexity and time consumption increase significantly
Solution Approach 1:
The patent segments the complex IC model into multiple neural network models, each trained on specific timescale data. This allows the system to handle different frequency ranges with specialized models rather than using a single comprehensive model, reducing overall complexity while maintaining accuracy across all timescales.
Solution Approach 2:
The patent performs preliminary training of neural network models on simulated data across various timescales before actual verification. This pre-training establishes accurate baseline predictions that can be quickly applied during verification, avoiding the need for complex real-time calculations while maintaining high precision.
2Measurement precision
If detailed modeling across all timescales is performed to capture transient spikes and dips, then measurement precision is improved, but loss of time increases
Solution Approach 1:
The neural network models are pre-trained on comprehensive simulated data covering all timescales including transient conditions. Once trained, these models can quickly predict waveforms during verification without requiring detailed real-time analysis, thus capturing transient spikes and dips accurately while significantly reducing verification time.
Solution Approach 2:
The patent creates simplified neural network copies of the complex IC behavior that can be evaluated much faster than the actual IC. These copies replicate the essential transient response characteristics, allowing rapid verification while maintaining detection accuracy for spikes and dips.
3Measurement precision
If synchronization across timescales is performed with detailed modeling, then measurement precision is improved, but device complexity and computational effort increase
Solution Approach 1:
The patent divides the synchronization task into multiple neural network models, each responsible for a specific timescale or frequency range. This segmentation simplifies the synchronization problem by handling each timescale independently with specialized models rather than attempting to synchronize all timescales simultaneously with a single complex model.
Solution Approach 2:
The patent transforms the complex multi-timescale synchronization problem into a more manageable form by using neural networks that can process and correlate data across different timescales through learned temporal relationships, effectively adding a dimensional transformation approach to the synchronization challenge.
Data Source
AI summary
A method includes obtaining a set of simulated signal values for an integrated circuit, providing the obtained set of simulated signal values to a machine learning model, and obtaining, from the machine learning model, one or more predicted output signal values of the integrated circuit. The method also includes comparing the predicted output signal values to actual output signal values of the integrated circuit to validate the integrated circuit.