CIM Macro Layout for Variable CNN Channel Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The fixed horizontal/vertical ratio of CIM macros in existing technologies is not optimized for varying input and output channels in convolutional neural networks, leading to inefficiencies in computation power and energy metrics.
Innovation Solution
A method to dynamically configure CIM macro arrangements based on the number of input and output channels of a designated convolutional layer, optimizing for latency, energy consumption, and utilization by using vertical, horizontal, or square arrangements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If multiple CIM macros are used to meet high computation requirements, then computation power is improved, but the fixed horizontal/vertical ratio causes inefficiency in computation power and energy metrics
Solution Approach 1:
The patent implements dynamic configuration of CIM macro arrangements by allowing the system to switch between different horizontal/vertical ratios based on the specific convolution operation requirements. The configuration manager dynamically selects optimal arrangements (e.g., 1:2, 2:1, 1:1 ratios) according to input/output channel dimensions, making the previously fixed ratio adaptable to varying computational needs, thereby improving both computation power utilization and energy efficiency.
2Device complexity
If fixed horizontal/vertical ratio is used for CIM macros, then device complexity is reduced, but adaptability to different input and output channels deteriorates
Solution Approach 1:
The patent creates a universal CIM macro arrangement system that can accommodate multiple convolution configurations through a configuration manager. This manager selects from pre-defined arrangement patterns (different horizontal/vertical ratios) based on the specific input and output channel dimensions. The system maintains low complexity by using a selection mechanism rather than complex dynamic reconfiguration, while achieving high adaptability through multiple supported arrangement patterns that can handle various CNN layer requirements.
3Ease of manufacture
If fixed CIM macro arrangement is used, then ease of manufacture is improved, but energy consumption optimization deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-defining multiple CIM macro arrangement patterns with different horizontal/vertical ratios during the design phase. These pre-configured patterns are stored and selected by the configuration manager based on the specific convolution operation requirements. This approach maintains manufacturing simplicity because the physical hardware structure remains fixed, while energy optimization is achieved through intelligent selection of the most efficient pre-defined arrangement for each computational task, avoiding the need for complex runtime reconfiguration.
Data Source
AI summary
A method and a non-transitory computer readable medium for CIM arrangement, and an electronic device applying the same are proposed. The method for CIM arrangement includes to obtain information of the number of CIM macros and information of the dimension of each of the CIM micros, to obtain information of the number of input channels and the number of output channels of a designated convolutional layer of a designate neural network, and to determine a CIM macro arrangement for arranging the CIM macros according to the number of the CIM macros, the dimension of each of the CIM macros, the number of the input channels and the number of the output channels of the designated convolutional layer of the designated neural network, for applying convolution operation to the input channels to generate the output channels.