Multi-Stream Neural Processing Under Memory Bandwidth Limits

The patent divides a single data stream into multiple parallel streams (e.g., multiple frequency bands or time segments) that are processed simultaneously by the same neural network weights. This segmentation allows the system to achieve better signal processing performance through parallel processing without increasing the size of the weight coefficients, thereby avoiding increased power consumption associated with transferring larger weight sets.

The patent creates multiple copies of the input data stream (augmented streams) that are processed in parallel using the same neural network weights. By copying and transforming the input data rather than copying the weights, the system improves performance through multiple processing paths while keeping the memory bandwidth and power requirements constant.

The patent segments the input data into multiple parallel streams that are processed simultaneously using the same weight coefficients. This approach increases the effective processing capacity and signal processing performance without requiring additional memory bandwidth for weight transfer, as the same weights are reused across all parallel streams.

The same set of neural network weights is made universal and reused across multiple parallel data streams. This multi-functionality allows the system to achieve improved signal processing performance through parallel processing while maintaining constant memory bandwidth requirements, as the weights serve multiple processing functions simultaneously.