Neural network computing method and system
By compiling asymmetric quantized neural network models into symmetric models and merging zeros to generate symmetric operations, the problem of high hardware costs is solved, achieving efficient neural network inference and reducing hardware costs.
CN117291275BActive Publication Date: 2026-06-09MEDIATEK INC
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MEDIATEK INC
- Filing Date
- 2022-06-24
- Publication Date
- 2026-06-09
AI Technical Summary
Technical Problem
Existing asymmetric quantization neural network models suffer from high hardware costs, high power consumption, and increased circuit area in hardware design, and the asymmetric quantization operation is also highly complex.
Method used
By using a compiler, asymmetric operations can be compiled into symmetric operations that include combined offset values. The zeros of the input and output can be merged to generate a symmetric neural network model, reducing computation and lowering hardware costs.
Benefits of technology
It enables efficient execution of neural network inference in fixed-point algorithms, reducing hardware cost and power consumption, while also reducing hardware area.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN117291275B_ABST
Abstract
The present disclosure provides neural network computing methods and systems that can reduce hardware cost. A neural network computing method of one embodiment can include receiving a neural network model comprising asymmetric operations, each asymmetric operation comprising one or more fixed-point operands that are derived from corresponding floating-point operands through asymmetric quantization; compiling, by a compiler, a given asymmetric operation of the neural network model into a symmetric operation comprising a combined offset value, wherein the combined offset value is a constant computed by the compiler by at least merging zero points of inputs and outputs of the given asymmetric operation; and generating a symmetric neural network model comprising the symmetric operation for execution in a fixed-point algorithm by inference hardware.
Need to check novelty before this filing date? Find Prior Art