DNN Weight Sparsity Balancing Across Processing Elements

The patent applies local quality by transforming the global sparsity property into local balance. Specifically, it redistributes sparse weights across PEs such that each PE receives a locally balanced portion with comparable numbers of zero and non-zero elements. This is achieved through algorithms that analyze the sparsity pattern and reassign weights to ensure uniform workload distribution, thereby resolving the synchronization bottleneck caused by unbalanced sparsity while preserving the computation efficiency gains from sparsity-based acceleration.

The patent addresses workload distribution by implementing local balance in sparsity exploitation. It redistributes weights across PEs to ensure each processor receives a balanced subset with comparable computational loads. This involves analyzing the sparsity patterns and reassigning weights through algorithms that minimize workload variance, thereby maintaining energy efficiency benefits while improving operational ease and load balancing across the parallel processing architecture.

The patent applies preliminary action by performing sparsity balancing and weight redistribution during the offline training or model preparation phase. The sparsity analysis, PE assignment, and weight reordering are all computed in advance before deployment. This preliminary processing creates a pre-optimized weight configuration that eliminates the need for complex runtime synchronization management, thereby reducing synchronization time without introducing significant overhead during actual DNN inference operations.