Method for processing activation data into multiple tiles using 3D convolution computation cores
By segmenting activation data into blocks and employing an adaptive 3D convolutional computation core and MAC array scheduling, the problem of low computational efficiency in convolutional neural networks is solved, achieving efficient resource utilization and energy consumption optimization, and adapting to various convolutional operation modes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BLACK SESAME TECH CO LTD
- Filing Date
- 2022-07-08
- Publication Date
- 2026-07-03
AI Technical Summary
The computational efficiency of convolutional neural networks in the current technology is low, especially when processing large-scale data. The resource utilization and energy consumption of hardware accelerators are high, making it difficult to meet the computational needs of complex neural networks.
The activation data is divided into multiple blocks, and an adaptive 3D convolutional computation core and MAC array scheduling method are used to dynamically adjust the MAC array to adapt to different modes, thereby improving computational efficiency.
It improves the computational efficiency of convolutional neural networks, reduces resource waste and energy consumption, supports multi-precision and sparse convolution operations, adapts to various tensors and kernel sizes, and achieves high-utilization computational performance gains.
Smart Images
Figure CN115169546B_ABST