Reconfigurable Parameter Circuit for Deep Learning Calculator Utilization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep learning inference units face memory bandwidth limitations and inefficient calculator utilization due to fixed parameters and network configurations, making them inflexible and slow for processing diverse CNN operations.
Innovation Solution
An information processing circuit with a product-sum circuit and a parameter value output circuit, where the latter is designed as a combinational circuit allowing configuration changes, enabling flexible operation across different deep learning tasks by optimizing parameter sets and network structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a common calculator is used to execute operations of multiple layers, then device complexity is reduced, but calculator utilization becomes inefficient due to memory bandwidth limitations
Solution Approach 1:
The patent segments the calculator into multiple independent calculators, each dedicated to a specific layer. This segmentation allows each calculator to operate independently without being bottlenecked by memory bandwidth, thereby improving calculator utilization while maintaining manageable device complexity through modular architecture.
Solution Approach 2:
Each calculator is designed with multi-functionality to handle various operations within its dedicated layer, including product-sum operations, activation functions, and parameter updates. This universality within each calculator unit improves overall system productivity without requiring excessive complexity at the system level.
2Manufacturing precision
If parameters and network configuration are fixed, then manufacturing precision is improved, but adaptability deteriorates for diverse CNN operations
Solution Approach 1:
The patent implements dynamic reconfigurability in the calculator units, allowing parameters and network configurations to be changed after manufacturing. This dynamic capability enables the same hardware to adapt to diverse CNN operations while maintaining manufacturing precision through standardized modular designs that can be programmed with different parameters.
Solution Approach 2:
The system allows parameter changes by providing interfaces to load different parameter sets into the calculators. This principle enables the same circuit configuration to perform different deep learning operations by simply changing the parameter values, thereby achieving adaptability without sacrificing manufacturing precision.
3Productivity
If calculators are provided for each layer, then calculator utilization is improved, but device complexity increases
Solution Approach 1:
The patent segments the deep learning network into multiple layers, with each layer having a dedicated calculator. This segmentation improves calculator utilization by eliminating memory bandwidth bottlenecks, while the modular nature of the segmentation keeps device complexity manageable through standardized reusable calculator units.
Solution Approach 2:
The patent resolves the complexity issue by organizing calculators in a layered dimensional structure rather than a flat monolithic architecture. Each calculator operates in its own computational dimension (layer), reducing inter-dependencies and managing complexity through spatial organization of computational resources.
Data Source
AI summary
The information processing circuit 10 performs operations on layers in deep learning, and includes a product sum circuit 11 which performs a product-sum operation using input data and parameter values, and a parameter value output circuit 12 which outputs the parameter values, wherein the parameter value output circuit 12 is composed of a combinational circuit, and includes a first parameter value output circuit 13 manufactured in a way that a circuit configuration cannot be changed and a second parameter value output circuit 14 manufactured in a way that allows a circuit configuration to be changed.