PIM Memory Controller Channel Remapping for Thermal Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing complexity of neural networks for deep learning applications leads to exponential computation requirements, and the performance of artificial intelligence systems is degraded due to limitations in data communication between separate memory and processor units, necessitating an integrated solution for improved data processing speed.
Innovation Solution
A memory system with a PIM device that integrates a processor and memory on a semiconductor chip, featuring a controller that remaps logical channel addresses to physical addresses based on temperature values to optimize data transmission and processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If separate memory and processor units are used, then device complexity is reduced, but data communication speed and AI performance are degraded
Solution Approach 1:
The patent merges memory and processor units into an integrated PIM device where processing elements are embedded within the memory structure. This allows data to be processed in-place without external data movement, resolving the contradiction by combining separate components to achieve both integration benefits and maintained complexity manageability through modular design.
Solution Approach 2:
The patent introduces a new architectural dimension by embedding processing elements within the memory hierarchy itself, creating a three-dimensional integrated structure. This dimensional transformation enables simultaneous access to storage and processing functions, improving data communication speed while managing system complexity through spatial organization.
2Reliability
If the number of neural network layers is increased, then AI performance is improved, but computation requirements increase exponentially
Solution Approach 1:
The patent segments the computational workload by distributing processing elements across multiple memory banks and channels. Each processing element handles specific neural network operations locally, dividing the exponential computation burden into manageable parallel segments that can be executed simultaneously, thus improving AI performance while controlling power requirements.
Solution Approach 2:
The patent enables continuous computation by maintaining data in memory during processing operations rather than requiring repeated data movement. Processing elements continuously operate on data residing in adjacent memory cells, eliminating idle time and ensuring continuous useful action throughout the computational pipeline, which improves efficiency for deep neural networks.
3Speed
If PIM device integrates processor and memory, then data processing speed is improved, but device complexity increases
Solution Approach 1:
The patent applies local quality by providing processing elements only where needed within specific memory banks, rather than uniformly across the entire device. Each memory bank contains processing elements optimized for its specific data access patterns, creating locally tailored functionality that improves processing speed while managing overall device complexity through selective integration.
Solution Approach 2:
The patent implements universal processing elements that can perform multiple operations (read, write, compute, and data movement) within the memory structure. These multi-functional elements reduce the need for separate dedicated circuits for each function, improving data processing speed while controlling device complexity through component versatility.
Data Source
AI summary
A memory system includes a plurality of memory dies respectively having at least one channel, a controller configured to control the plurality of memory dies, and a base die configured for interfacing signal and data transmissions between the plurality of memory dies and the controller. The controller is configured to remap a logical channel address of the most frequently used channel to a physical channel address of a channel having a lowest temperature value to transmit the remapped physical channel address to the base die.


