System-Level Yield Estimation via Statistical Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional yield models fail to accurately analyze the impact of process variability on electronic systems, particularly in nanometer technology, leading to functional and parametric yield loss, and do not effectively capture the effects of frequency scaling in embedded real-time systems.
Innovation Solution
A method for determining system-level yield loss by propagating statistical properties of individual components' performance variables, such as delay and energy, to the system level, while considering application information and correlations, using bivariate probability density functions to estimate yield loss and energy consumption, allowing for run-time trade-offs and dynamic frequency scaling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional yield models are used to analyze electronic systems, then the analysis process is simple, but the accuracy of yield estimation is insufficient
Solution Approach 1:
The patent segments the yield analysis into component-level and system-level analyses. Statistical properties are first obtained for individual components, then propagated to the system level through mathematical operations (convolution for energy, maximum for delay). This segmentation allows accurate system-level yield estimation without requiring exhaustive system-level simulations, thus improving accuracy while managing complexity.
Solution Approach 2:
The patent introduces statistical properties (mean, standard deviation, probability density functions) as intermediary representations between component characteristics and system-level yield metrics. These statistical intermediaries enable accurate yield propagation through mathematical operations, bridging the gap between component-level data and system-level predictions without direct complex simulations.
2Reliability
If frequency binning is applied to increase functional yield, then more chips become operationally functional, but timing parametric yield deteriorates due to missed application deadlines
Solution Approach 1:
The patent introduces dynamic frequency scaling as a runtime mechanism to adjust operating frequency based on actual component performance. Instead of static frequency binning, the system dynamically adapts frequency to meet application deadlines while maximizing functional utilization. This dynamic approach resolves the contradiction by allowing frequency adjustment that maintains both functional yield and timing parametric yield.
Solution Approach 2:
The patent changes the operating frequency parameter at runtime based on measured component performance and application deadline requirements. By dynamically adjusting this critical parameter, the system can maintain timing parametric yield (meeting deadlines) while maximizing functional yield (utilizing more chips), thus resolving the trade-off between the two yield types.
3Measurement precision
If deterministic and predictable timing and energy consumption is assumed at system level, then design estimation is simplified, but accuracy deteriorates due to process variability
Solution Approach 1:
The patent segments the statistical propagation into distinct mathematical operations for different parameters: convolution for energy consumption and maximum operation for delay. This segmentation allows accurate handling of process variability through appropriate statistical methods for each parameter type, improving estimation accuracy while managing the complexity through structured mathematical approaches.
Solution Approach 2:
The patent uses statistical properties (probability density functions, mean, standard deviation) as intermediary representations to propagate variability effects from components to system level. These statistical intermediaries enable accurate prediction of timing and energy consumption under process variability without requiring complex simulations, thus improving accuracy while managing computational complexity.
4Manufacturing precision
If all system components must meet the same performance requirements, then timing yield is maximized, but energy consumption increases and area grows
Solution Approach 1:
The patent applies local quality by allowing different components to have different performance characteristics based on their actual measured properties. Instead of requiring all components to meet uniform performance requirements, the system adapts operating parameters (frequency, voltage) locally to each component's capabilities. This resolves the contradiction by maintaining timing yield through adaptive operations while reducing energy consumption and area through component-specific optimization.
Solution Approach 2:
The patent introduces dynamic adaptation of operating parameters based on individual component performance measurements. Components can operate at different frequencies and voltages according to their actual characteristics rather than uniform specifications. This dynamic approach maintains timing yield by ensuring each component operates within acceptable parameters while minimizing energy consumption through optimized operating points for each component.
Data Source
AI summary
One inventive aspect relates to a method of determining an estimate of system-level yield loss for an electronic system comprising individual components subject to manufacturing process variability leading to manufacturing defects. The method comprises obtaining a description of the composition of the electronic system in terms of which individual components are used. The method further comprises obtaining statistical properties of the performance of individual components of the electronic system with respect to first and second performance variables, e.g. energy consumption and delay, the statistical properties including correlation information of the first and second performance variables. The method further comprises obtaining information about execution of an application on the system, e.g. a number of accesses of a component by an application. The method further comprises propagating the statistical properties of the first and second performance variables of the individual components to the electronic system so that the correlations between the first and second performance variables are preserved, the propagating taking into account the application information.