Network Devices Offload Neural Network Computations

The neural network computation is segmented into two parts: a subset of computations is performed by network devices on network packets, while the remaining computations are performed by servers. This segmentation allows the network devices to offload computational tasks to servers, reducing server load and improving overall system productivity without requiring network devices to perform all computations alone.

Network devices act as intermediaries between clients and servers by performing partial neural network computations on network packets and encoding the results in packet headers. This intermediary role allows servers to receive pre-processed information and reduce their computational workload, effectively mediating the computation distribution without requiring servers to handle all processing themselves.

The patent replaces traditional hardware-based lookup tables with neural network computations performed by programmable switching chips. Instead of using large memory structures for packet classification, the system uses computational models that can be configured through software, reducing the need for expensive hardware resources while maintaining or improving classification capabilities.

The system changes the approach to packet classification from static lookup tables to dynamic neural network computations. By parameterizing the classification function through learnable models rather than fixed hardware entries, the system achieves adaptability without proportionally increasing hardware cost, as the computational models can be trained and configured through software updates.