Neural Network Computation With Fixed-Point Data for Lower Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network operations based on floating-point data require high memory and energy consumption due to strict memory requirements, leading to high costs.
Innovation Solution
Utilizing fixed-point data for neural network computations by converting floating-point data to fixed-point data using preset conversion parameters, reducing memory and operation amounts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point data is used for neural network operations, then computation precision is maintained, but memory requirements and energy consumption increase significantly
Solution Approach 1:
The patent changes the data representation parameter from floating-point to fixed-point format. This parameter change reduces the memory footprint and energy consumption while maintaining acceptable computational precision for neural network operations, directly resolving the contradiction between precision and memory requirements
Solution Approach 2:
The patent employs fixed-point arithmetic which can be implemented using simpler, less expensive hardware components compared to floating-point units. This allows the system to use cheaper computational resources while achieving the same functional outcome, addressing the memory and energy consumption issues
2Measurement precision
If floating-point data is used for neural network operations, then computation precision is maintained, but energy consumption increases
Solution Approach 1:
By changing the data type parameter from floating-point to fixed-point, the patent reduces the energy consumption of neural network operations. Fixed-point arithmetic requires less energy for storage and computation while maintaining sufficient precision for the application domain
Solution Approach 2:
The patent uses fixed-point arithmetic which can be implemented with simpler, lower-power hardware circuits. This replaces expensive floating-point units with more energy-efficient alternatives, directly reducing the energy consumption of the neural network computation device
3Quantity of substance
If fixed-point data is used for neural network operations, then memory requirements and energy consumption are reduced, but computational complexity increases due to conversion operations
Solution Approach 1:
The patent performs floating-point to fixed-point conversion in advance, before the actual neural network computations. This preliminary action converts all input data and weights to fixed-point format once, and then the subsequent computations can proceed using simple fixed-point arithmetic without repeated conversions, reducing overall system complexity
Solution Approach 2:
The patent introduces a conversion module as an intermediary component that handles the transformation between floating-point and fixed-point representations. This dedicated intermediary handles the complexity of data conversion separately from the main computation pipeline, allowing the core computational operations to remain simple and efficient
Data Source
AI summary
The disclosure provides a neural network computation device, a neural network computation method, and related products, which are applied to a neural network chip. The neural network chip is arranged on a board card. The board includes a storage component configured to store data, an interface means configured to realize data transfer between the neural network chip and an external device; and a control component configured to monitor a status of the neural network chip. The device includes: an operation unit, a controller unit, and a storage unit. The device is used for performing artificial neural network training operations. The neural network training operations include neural network multi-layer training operations. The technical solution provided by the present disclosure has the technical effects of low cost and low energy consumption.