Array Data Scaling Using Local Maxima for FP8 Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing techniques do not effectively reduce memory access overheads during data format conversion, particularly in the context of machine learning operations, leading to inefficiencies in data representation and processing.
Innovation Solution
A data scaling system that identifies the largest value in each portion of an array and uses this value to sequentially scale the data within each portion, employing both local and global scaling factors to convert data formats efficiently, such as from FP32 to FP8, thereby reducing memory access overheads.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data is processed using traditional memory access methods during data format conversion, then data processing can be performed, but memory access overheads are high leading to processing inefficiency
Solution Approach 1:
The patent divides the data array into multiple portions and processes each portion independently to identify local maximum values. This segmentation allows parallel processing of different data portions, reducing overall memory access overhead and improving processing efficiency while maintaining data integrity across the entire array.
Solution Approach 2:
The patent performs preliminary identification of maximum values in each data portion before executing the full data processing operation. By pre-calculating scaling factors based on local maximums, the system avoids repeated memory accesses during the main processing phase, thereby reducing memory access overhead and improving productivity.
2Productivity
If data is scaled using global maximum values only, then data format conversion can be performed, but processing efficiency is reduced due to sequential processing requirements
Solution Approach 1:
The patent applies different scaling approaches to different portions of the data array. Each portion is scaled using its own local maximum value, allowing independent and potentially parallel processing of each portion. This local quality approach improves processing efficiency while the systematic method for determining local maximums keeps the overall complexity manageable.
Solution Approach 2:
The patent transitions from a single global scaling dimension to multiple local scaling dimensions by identifying maximum values in each data portion separately. This dimensional change allows parallel processing paths for different portions, improving efficiency while the structured approach to multi-dimensional scaling maintains acceptable complexity levels.
3Loss of time
If data portions are processed independently with local maximum values, then memory access overhead is reduced, but data representation accuracy may be compromised
Solution Approach 1:
The patent incorporates feedback mechanisms where the identified local maximum values are used to determine scaling factors that are then applied back to the original data portions. This feedback loop ensures that each portion is scaled appropriately based on its own characteristics, maintaining data representation accuracy while the localized nature of the feedback reduces memory access overhead compared to global processing.
Data Source
AI summary
Apparatuses, systems, and techniques to scale values. In at least one embodiment, a processor comprising one or more circuits to cause a largest value of each portion of two or more portions of an array to be identified and to use the largest value of each portion to scale one or more values within each portion sequentially.