Branch Predictor Bias-Free Filtering for Deeper Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current branch predictors, especially those with moderate hardware budgets, are limited in predicting branch directions due to restricted correlation capabilities over larger distances, leading to increased access latencies and power consumption when attempting to track more branches.
Innovation Solution
A branch predictor apparatus that filters biased conditional branches and tracks only the latest occurrences of non-biased branches, allowing for deeper correlation analysis and improved prediction accuracy by using a biased branch filter and correlation element within a branch history buffer, along with a neural-based perceptron predictor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If branch prediction hardware budget is increased to track more branches over larger distances, then prediction accuracy is improved, but access latency and power consumption increase
Solution Approach 1:
The patent extracts and removes biased branches from the branch history buffer, keeping only non-biased branches. This selective extraction reduces the number of entries that need to be tracked and correlated, thereby reducing hardware budget, access latency, and power consumption while maintaining prediction accuracy for the most relevant branches.
Solution Approach 2:
The patent applies different treatment to different types of branches: biased branches are filtered out while non-biased branches are retained and tracked. This local differentiation optimizes resource allocation by focusing hardware resources on tracking only the branches that provide meaningful correlation signals.
2Measurement precision
If branch prediction hardware budget is increased to track more branches over larger distances, then prediction accuracy is improved, but access latency increases
Solution Approach 1:
By extracting only non-biased branches from the history buffer, the patent reduces the total number of entries that need to be searched and correlated. This smaller, filtered set reduces access latency while maintaining the ability to capture long-distance correlations that improve prediction accuracy.
3Use of energy by moving object
If moderate hardware budget (32-64 KB) is used in branch predictor, then power consumption is reduced, but correlation capability over large distances is limited
Solution Approach 1:
The patent extracts only the most valuable branch information (non-biased branches) from the instruction stream, creating a compact but highly effective history buffer. This selective filtering allows moderate hardware budgets to achieve better prediction accuracy by focusing resources on branches that provide meaningful correlation signals over long distances.
4Measurement precision
If larger data structures are used to track more branches, then prediction accuracy is improved, but training time for neural-based perceptron predictors increases
Solution Approach 1:
By extracting only non-biased branches into the history buffer, the patent creates a smaller, more focused training dataset for neural-based perceptron predictors. This reduced dataset requires less training time while maintaining the ability to capture long-distance correlations, as biased branches do not provide useful correlation signals.
Data Source
AI summary
An apparatus and method for filtering biased conditional branches in a branch predictor in favor of non-biased conditional branches are disclosed. Biased conditional branches, which are consistently skewed toward one direction or outcome, are filtered such that an increased number of non-biased conditional branches which resolve in both directions may be considered. As a result, more useful branches may be captured over larger distances, thereby providing correlations deeper in a global history. In addition, by tracking only the latest occurrences of non-biased conditional branches using a recency stack structure, even more distant branch correlations may be made.