PIM Memory Device Pruning Neural Network Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the size of data increases in machine learning, the computational load for neural network operations increases, leading to decreased computational speed and efficiency, which is not effectively addressed by existing memory devices.
Innovation Solution
A memory device with a processing-in-memory (PIM) chip that performs pruning operations on activation and weight data, reducing the computational load by omitting operations with small influence on processing results, and providing pruned data to the host device for neural network operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the host device performs pruning operations on activation data and weight data, then the processing accuracy is improved, but the computational load on the host device increases and computational speed decreases
Solution Approach 1:
The patent extracts the pruning operation function from the host device and relocates it to the memory device. The memory device independently performs pruning on activation data and weight data, separating this computationally intensive task from the host device's neural network processing workload, thereby improving host computational speed while maintaining processing accuracy
Solution Approach 2:
The memory device acts as an intermediary between data storage and host processing. It performs pruning operations as an intermediate processing step before providing pruned data to the host device, reducing the computational burden on the host while ensuring accurate processing results
2Quantity of substance
If the data size increases for machine learning, then the model capacity and accuracy are improved, but the computational load increases and system efficiency decreases
Solution Approach 1:
The patent applies local quality by performing pruning operations selectively on specific data elements (activation data and weight data) rather than processing all data uniformly. The memory device identifies and removes less important data elements locally, maintaining the essential information while reducing overall data volume and computational requirements
Solution Approach 2:
The patent changes the data parameters by transforming full-precision activation data and weight data into pruned data with reduced precision. This parameter change reduces the computational load for subsequent neural network operations while maintaining adequate model accuracy
3Reliability
If the host device performs all neural network operations, then the processing completeness is improved, but the device complexity and computational burden increase
Solution Approach 1:
The patent segments the neural network processing workflow into distinct functional parts: the memory device handles data preparation and pruning operations, while the host device focuses on core neural network computations. This segmentation distributes computational burden across multiple components, reducing the complexity of any single device while maintaining complete processing functionality
Data Source
AI summary
A method of operating a memory device which is configured to communicate with a host device, and which includes a memory cell array and a processing-in-memory (PIM) chip, including loading activation data and weight data from the memory cell array, by the PIM chip; generating pruned activation data by performing a first pruning operation on the activation data, by the PIM chip; providing the pruned activation data to the host device, by the PIM chip; generating pruned weight data by performing a second pruning operation on the weight data, by the PIM chip; providing the pruned weight data to the host device, by the PIM chip; and storing output data corresponding to a neural network operation performed based on the pruned activation data and the pruned weight data, by the memory cell array.