Matrix calculation method, device and related equipment
By adjusting the target bit width area of the target element to 0 in the matrix calculation, the problem of difficulty in reducing the power consumption of single data format matrix calculation in the existing technology is solved, and the power consumption is reduced and the efficiency is improved while meeting the accuracy requirements.
Patent Information
- Application Number
- CN202411075977.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-07
- Publication Date
- 2025-09-09
- Estimated Expiration
- 2044-08-07
AI Technical Summary
When a processing unit performs matrix calculations in a single data format, existing technologies have difficulty in effectively reducing power consumption, especially when the data format is fixed and the computational complexity cannot be reduced by quantizing or reducing the element bit width.
The complexity of matrix calculation is reduced by obtaining a target bit-width region of a target element in a matrix calculation and adjusting the value of the region to 0. The specific method includes generating a mask and using the mask to adjust the value of the target bit-width region to 0.
On the premise of meeting the calculation accuracy requirements, the power consumption of matrix calculation is reduced and the calculation efficiency is improved.