Crossbar Neural Network Pruning via Row-Column Sparsity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network accelerators based on crossbar architectures face inefficiencies due to high resource consumption and hardware waste, as they primarily utilize sparsity in the column direction of weight matrices, neglecting sparsity in the row direction and leading to idle cells.
Innovation Solution
A pruning method that employs structured pruning across multiple dimensions, including block-based, crossbar row, and crossbar column pruning, dynamically adjusting thresholds to optimize weight matrix mapping onto crossbars, utilizing XRST and XCST modules for efficient sparsity handling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional column-based pruning is applied to reduce model size, then the compression ratio is limited, but hardware resources are wasted due to idle cells in crossbar architecture
Solution Approach 1:
The patent transitions from traditional column-based pruning to two-dimensional row-column joint pruning. By introducing row-direction pruning alongside column-direction pruning, the method fully utilizes sparsity in both dimensions of the weight matrix, thereby improving compression ratio while avoiding idle cells in the crossbar architecture.
Solution Approach 2:
The patent divides the weight matrix into multiple blocks and applies pruning operations to each block independently. This segmentation allows for fine-grained control over the pruning process, enabling the method to achieve high compression ratios while maintaining compatibility with the crossbar architecture's resource allocation.
2Reliability
If crossbar resources are allocated to accommodate full weight matrices, then hardware resources are over-provisioned, but resource utilization efficiency decreases
Solution Approach 1:
The patent dynamically adjusts pruning thresholds based on the specific characteristics of each weight matrix and the corresponding crossbar resource capacity. By changing the pruning parameters adaptively, the method achieves optimal balance between maintaining computation accuracy and improving resource utilization efficiency.
Solution Approach 2:
The patent implements a dynamic pruning strategy where the pruning ratio and threshold are adjusted according to the importance of different weight elements and the available crossbar resources. This dynamic approach allows the system to maintain high accuracy while maximizing resource utilization by pruning less important weights more aggressively.
Data Source
AI summary
The present disclosure relates to a pruning method based on a crossbar architecture and a system thereof, wherein the method comprises the steps of: carrying out structured pruning on a weight matrix to be processed at the current layer of a neural network without considering architecture factors to obtain a first weight matrix, wherein the weight matrix is the weight matrix of the neural network; determining the pruning scheme according to the size of the first weight matrix and the size of the crossbar, wherein the pruning scheme comprises: block-based structured pruning, crossbar row pruning and crossbar column pruning; and pruning the first weight matrix according to the pruning scheme to obtain a second weight matrix, wherein the second weight matrix is mapped to the crossbar array in an accelerator, and the neural network is accelerated by the accelerator.