Floating Point Accumulation Circuit Using Exponent Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing floating point number accumulation methods in computers require large layout areas and suffer from precision issues, especially in applications like neural networks, where conversion to fixed point numbers is necessary, leading to inconsistencies in operations.
Innovation Solution
An accumulation apparatus and method that identifies a base exponent among a cluster of floating point numbers, filters a subset for accumulation based on this exponent, and performs accumulation without significant precision loss, using an identification unit, filtering unit, and addition unit to align exponents and shift mantissas accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a multi-level accumulation tree is used to accumulate floating point numbers, then the accumulation can be completed, but the layout area becomes large and the precision is not ideal
Solution Approach 1:
The patent divides the floating point numbers into multiple clusters based on their exponent ranges. Each cluster is processed separately with its own accumulation tree, rather than using a single large accumulation tree for all numbers. This segmentation reduces the layout area while maintaining precision within each cluster.
Solution Approach 2:
The patent introduces a new dimension of organization by grouping floating point numbers according to their exponent values into different clusters. This dimensional organization allows for more efficient resource utilization and reduces the overall layout area compared to a flat single-tree structure.
2Adaptability or versatility
If floating point number accumulation is performed and then converted to fixed point number for Wallace tree multiplication, then the operations can proceed, but the consistency of operation is compromised
Solution Approach 1:
The patent performs preliminary clustering of floating point numbers by exponent range before the accumulation operation. This preliminary organization ensures that numbers within each cluster have compatible exponent values, allowing for more accurate conversion to fixed point format and maintaining operational consistency throughout the computation pipeline.
3Productivity
If a large accumulation tree is used to handle all floating point numbers, then complete accumulation is achieved, but the circuit complexity and power consumption increase
Solution Approach 1:
The patent segments the accumulation task into multiple smaller accumulation trees, each handling a specific cluster of floating point numbers with similar exponent ranges. This segmentation reduces the complexity and power consumption of each individual tree while maintaining overall accumulation throughput through parallel processing of multiple clusters.
Solution Approach 2:
The patent dynamically organizes floating point numbers into clusters based on their exponent values, allowing the system to adaptively adjust the structure and size of accumulation trees according to the actual data distribution, thereby optimizing circuit complexity and power consumption for different input scenarios.
Data Source
AI summary
An accumulation apparatus according to an embodiment accumulates a plurality of floating point numbers in an identification cluster. A base exponent is identified, and, then, an accumulation cluster is filtered according to the base exponent, and floating point numbers in the accumulation cluster are accumulated. A small circuit area, low power consumption, and high precision can be achieved.