Multi-Precision Data Truncation for Memory Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern computing workloads, particularly those involving large data processing like machine learning and artificial intelligence, face memory retrieval bottlenecks due to the speed at which data is retrieved from memory, which can decrease overall computational throughput.
Innovation Solution
Implementing a system that truncates data into multiple precision levels, allowing for separate storage and retrieval of data portions at different precision levels, including full and lower precision, with the option to add adjustment values to maintain accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data is retrieved from memory at full precision, then measurement precision is maintained, but memory throughput decreases and power consumption increases
Solution Approach 1:
The patent segments data into multiple precision levels (e.g., full precision and truncated precision portions). Different portions of data are stored at different precision levels, allowing the system to retrieve only the necessary precision level for each computational task, thereby improving memory throughput while maintaining required precision.
Solution Approach 2:
The patent changes the precision parameter of data storage and retrieval dynamically. By adjusting the precision level based on computational requirements, the system can optimize between memory throughput and power consumption while maintaining sufficient measurement precision for the task at hand.
2Measurement precision
If data is retrieved from memory at full precision, then measurement precision is maintained, but power consumption increases
Solution Approach 1:
The patent segments data into multiple precision levels, allowing selective retrieval of data at appropriate precision levels. This reduces the total amount of data transferred and processed, thereby reducing power consumption while maintaining required precision for computational accuracy.
Solution Approach 2:
The patent dynamically changes the precision parameter of data retrieval based on computational requirements. By using lower precision when full precision is not needed, the system reduces power consumption associated with data transfer and processing while maintaining sufficient accuracy.
3Productivity
If data is truncated to lower precision levels, then memory throughput increases and power consumption decreases, but measurement precision decreases
Solution Approach 1:
The patent segments data into full precision and truncated precision portions stored separately. The system can retrieve only the truncated portions for tasks that don't require full precision, improving memory throughput, while full precision data remains available when higher accuracy is needed.
Solution Approach 2:
The patent applies partial precision retrieval by fetching only the necessary precision level for each task. This partial action approach retrieves less data than full precision would require, improving throughput while maintaining sufficient precision through selective truncation.
4Use of energy by moving object
If data is truncated to lower precision levels, then power consumption decreases, but measurement precision decreases
Solution Approach 1:
The patent segments data into multiple precision levels, enabling the system to choose appropriate precision levels based on task requirements. This segmentation allows power consumption to be reduced by using lower precision when sufficient, while maintaining the option to use full precision when accuracy is critical.
Data Source
AI summary
Systems and methods for data truncation are disclosed. A computing system may a storage medium and a processor. The processor may be configured to identify first data of a first data structure; identify a first portion of the first data and a second portion of the first data; store the first portion of the first data as a second data structure in the storage medium, and the second portion of the first data as a third data structure in the storage medium; identify a request for the first data; based on the request, retrieve from the storage medium the second data structure including the first portion of the first data; identify a value associated with the second portion; generate second data based on the first portion and the value; and perform a task based on the second data.