Data processing apparatus, processor, board card, and data processing method
By integrating particle quantization/dequantization into vector operations, adopting the MX quantization format, and optimizing the scaling factor, the hardware compatibility and computational efficiency issues of neural network models are resolved, thereby improving computational accuracy and resource utilization efficiency.
Patent Information
- Application Number
- PCT/CN2025/090096
- 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
In existing technologies, neural network models suffer from floating-point format migration issues during training and inference, resulting in poor hardware compatibility. Furthermore, the lack of effective quantization/dequantization processing schemes in vector operations affects computational efficiency and accuracy.
A data processing apparatus and method are provided, which reduces storage space and bandwidth waste by integrating particle quantization/dequantization processing during vector operations, adopting the MX quantization format, and optimizing the calculation and compression of scaling factors by extending the MX quantization scheme.
It improves the computational efficiency and accuracy of neural network models, reduces the amount of computation and computational resource requirements, and achieves data format compatibility across hardware platforms.