Memory Controller Data Separation for Differential Refresh Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing amount of memory required for machine learning tasks leads to high power consumption and greenhouse gas emissions, necessitating a technology to reduce power consumption in memory systems.
Innovation Solution
A memory system with separate storage regions for integer and decimal data parts, utilizing different refresh rates and storage characteristics to optimize power usage, including a data separator, address generator, and command generator to manage these regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a large amount of memory is used to store data for machine learning, then the data storage capacity increases, but the power consumption increases
Solution Approach 1:
The memory is divided into two distinct storage regions: a first storage region for integer data and a second storage region for decimal data. This segmentation allows different refresh rates to be applied to each region, reducing overall power consumption while maintaining necessary data accuracy for machine learning operations.
Solution Approach 2:
Different storage regions are assigned different refresh rates based on their specific data requirements. The first storage region (integer data) uses a higher refresh rate to maintain precision, while the second storage region (decimal data) uses a lower refresh rate, optimizing power consumption locally for each data type's needs.
2Measurement precision
If the refresh rate for maintaining data accuracy is increased, then the data precision is improved, but the power consumption increases
Solution Approach 1:
The data is segmented into integer and decimal portions, stored in separate regions, allowing differentiated refresh strategies. Integer data requiring high precision maintains a higher refresh rate, while decimal data can tolerate a lower refresh rate, reducing overall power consumption while preserving critical precision requirements.
Solution Approach 2:
Each storage region is configured with a refresh rate appropriate to its data type's precision requirements. The first storage region uses a higher refresh rate for integer data that demands greater accuracy, while the second storage region uses a lower refresh rate for decimal data, optimizing the local quality of power consumption to match local precision needs.
Data Source
AI summary
A memory controller includes: a data separator configured to separate host write data into upper data and lower data; an address generator configured to generate a first address and a second address based on a host address; a command generator configured to generate one or more first commands for writing the upper data into a first storage region that is selected based on the first address in a memory, and one or more second commands for writing the lower data into a second storage region that is selected based on the second address in the memory; and a control block configured to control the address generator and the command generator to make a difference in power consumption between the first storage region and the second storage region.