Data processing apparatus, processor, board, and data processing method
By incorporating scaling factors and supporting granular quantization in convolution operations, the problem of floating-point format migration between different AI vendors is solved, improving the efficiency and accuracy of neural network operations, reducing resource overhead, and adapting to the mixed operation needs of multiple data types.
Patent Information
- Application Number
- PCT/CN2025/090097
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-09-25
- Filing Date
- 2025-04-21
- Publication Date
- 2025-12-26
AI Technical Summary
Migration of existing floating-point formats between different AI vendors is difficult, and traditional quantization formats suffer from high resource consumption and severe accuracy loss in neural network operations, especially in convolution operations where they are difficult to be compatible with MX quantization formats.
A convolution operation scheme compatible with MX quantization format is provided. By fusing the scaling coefficients of input neurons and weights in the convolution operation, it supports granular quantization, performs data type conversion and compression when necessary, simplifies dequantization processing, and supports mixed operations of multiple data types.
It improves the efficiency and accuracy of neural network operations, reduces the overhead of storage and computing resources, supports mixed operations of multiple data types, and adapts to different data characteristics and accuracy requirements.