Block Floating-Point Compression for Large Exponent Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing methods using floating-point formats face inefficiencies in operations involving large exponential differences between data sets, leading to increased computational complexity and precision loss due to the need for separate exponent and mantissa handling.
Innovation Solution
Implementing a block floating-point format where shared exponents are represented by an exponent identifier field, allowing for separate processing of exponents and mantissas based on exponential differences, and using a lazy update mechanism for accumulation when differences are significant, thereby optimizing operations and reducing bit usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate exponent and mantissa handling is used in floating-point operations, then precision is maintained, but computational complexity increases
Solution Approach 1:
The floating-point number is segmented into exponent field and mantissa field, allowing independent processing of each component. This segmentation enables optimized handling where exponents are processed separately from mantissas, reducing overall computational complexity while maintaining precision through dedicated processing paths for each segment.
Solution Approach 2:
The patent implements dynamic processing by detecting exponential differences between operands and adapting the operation scheme accordingly. When exponential difference exceeds a threshold, lazy update mechanisms are activated; otherwise, separate exponent and mantissa processing is used, creating a dynamic system that adjusts computational complexity based on input characteristics.
2Adaptability or versatility
If more bits are allocated to exponents, then range of numbers increases, but bits available for mantissas decreases
Solution Approach 1:
The patent employs dynamic bit allocation where the number of bits used for exponents versus mantissas is not fixed but adapts based on the exponential difference between operands. This dynamic approach allows the system to optimize the balance between range and precision for each specific operation, rather than being constrained by a static allocation scheme.
Solution Approach 2:
The system changes parameters (bit allocation, processing mode) based on the detected exponential difference. When the difference is small, more bits are effectively allocated to mantissas for precision; when the difference is large, the system switches to lazy update modes that reduce the effective precision requirements, thereby optimizing the range-precision tradeoff dynamically.
3Productivity
If lazy update mechanism is used for accumulation, then computational overhead is reduced, but precision may be affected
Solution Approach 1:
The lazy update mechanism is dynamically activated or deactivated based on whether the exponential difference exceeds a predetermined threshold. This dynamic control allows the system to switch between high-efficiency lazy update mode and high-precision separate processing mode, optimizing the balance between productivity and precision based on the specific computational context.
Solution Approach 2:
The system changes the processing parameter (update frequency, precision level) based on the exponential difference parameter. When exponential difference is large, lazy update with lower precision is acceptable and improves efficiency; when exponential difference is small, higher precision processing is applied to maintain accuracy, thus adaptively adjusting parameters to resolve the contradiction.
Data Source
AI summary
An electronic device that compresses data and an operating method thereof are provided. The electronic device includes a processor configured to express each of a plurality of data according to a floating-point format that includes a sign field, an exponent identifier field, and a mantissa field, wherein an exponent identifier field included in each of the plurality of data includes a bit value that represents any one of a plurality of exponents.