Descriptor Synchronization for Neural Network Data Processing

The patent applies preliminary action by pre-defining data access patterns and address calculation rules before actual data processing. The system pre-configures descriptor information that contains data shape, stride, and offset parameters, so that during runtime, processors can directly use these pre-calculated values without performing complex address determination operations, thus improving processing efficiency while maintaining data access control

The patent introduces an intermediary mechanism in the form of a descriptor structure that mediates between the processor and the actual data. This descriptor contains pre-computed address information and data metadata, allowing the processor to access data through simplified instructions while the intermediary handles the complex address calculation and data layout management, thereby resolving the contradiction between controlled access and processing efficiency

The patent applies segmentation by dividing the data synchronization process into independent descriptor-level operations. Instead of synchronizing entire large datasets, the system segments data into tensor elements with associated descriptors that can be synchronized independently. This allows processors to synchronize only the necessary descriptor information rather than all data, significantly reducing synchronization overhead while maintaining the ability to handle complex multi-dimensional AI algorithms

The patent uses copying by creating and transmitting lightweight descriptor copies between processors rather than copying actual data. The descriptor contains metadata about the data (shape, stride, offset) that can be replicated without the burden of the actual data payload. This copying approach allows complex AI algorithms to be supported across multiple processors while minimizing the time loss associated with data synchronization