Statistical Data Bus Inversion Circuit for Low Latency Energy Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data bus inversion (DBI) techniques face high latency issues due to the use of population count circuits, which limits their application in on-chip buses where low latency is critical, especially in large network-on-chip buses where port arbitration is not known in advance, resulting in only 50-75% effective energy savings.
Innovation Solution
A low latency statistical DBI circuit using adjacent pattern prediction logic to determine whether data should be inverted, reducing latency and overhead while maintaining energy savings by approximating DBI functionality with simpler AND-OR pattern detectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If population count circuit is used for DBI, then energy savings are achieved, but latency increases significantly
Solution Approach 1:
The patent segments the DBI decision-making process into two parts: a fast statistical approximation using simple logic circuits for most cases, and a more accurate population count method only when needed. This segmentation allows the system to achieve energy savings with reduced latency by avoiding the expensive population count circuit in the critical data path.
Solution Approach 2:
The patent uses cheap, simple logic circuits (AND-OR pattern detectors) that provide approximate DBI functionality for most cases. These simple circuits have low latency and can be quickly evaluated, serving as a disposable approximation that doesn't require the complex, high-latency population count circuit.
2Loss of energy
If population count circuit is used for DBI, then toggle reduction is accurately determined, but circuit complexity and overhead increase
Solution Approach 1:
The patent segments the DBI implementation into a primary fast path using simple logic circuits and a secondary path for edge cases. The simple AND-OR pattern detectors provide the majority of DBI functionality with minimal complexity, while population count is reserved only for cases where the approximation is insufficient.
Solution Approach 2:
The patent replaces the expensive, complex population count circuit with cheap, simple logic circuits that approximate DBI functionality. These simple circuits consume fewer resources and have lower complexity, providing sufficient accuracy for most practical cases without the overhead of full population counting.
3Loss of energy
If DBI is applied in network-on-chip with unknown arbitration, then energy savings are reduced to 50-75%, but latency remains high with population count
Solution Approach 1:
The patent performs preliminary DBI approximation using simple logic circuits before the arbitration outcome is known. This preliminary action provides a fast decision that can be applied immediately, avoiding the need to wait for arbitration completion before making the DBI decision, thus reducing latency in network-on-chip applications.
Solution Approach 2:
The patent makes the DBI decision dynamic and adaptive, using the fast approximation method that can respond to changing data patterns in real-time without being constrained by the arbitration process. This dynamic approach allows the system to achieve energy savings even when arbitration outcomes are unpredictable.
Data Source
AI summary
In accordance with embodiments disclosed herein, there is provided systems and methods for low latency statistical data bus inversion (DBI) for energy reduction. A transmitting component includes a transmitter and a statistical DBI circuit. The statistical DBI circuit is to receive current data to be transmitted on a data bus and is to store previous data transmitted on the data bus. The statistical DBI circuit includes inverting logic to invert bits of the current data before transmission in response to a control signal. The statistical DBI circuit includes adjacent pattern prediction logic to receive a difference vector including a comparison of the previous data and the current data, determine whether the difference vector includes a pattern predicting transmission of the current data with toggle is more efficient than without toggle, and output the control signal in the first state indicating the pattern was detected.