In-Network Computing

In-network computing is the architectural paradigm that moves computation from end-host servers into the network fabric — switches, routers, programmable NICs, and smart NICs — to reduce data movement and exploit the data locality principle at the network layer. It is a direct response to the communication-bound regime: when the cost of transferring data across a datacenter network exceeds the cost of processing it, the optimal architecture is one that processes data where it already flows.\n\nThe enabling technology is programmable network hardware: P4-programmable switches, RDMA-capable NICs, and SmartNICs with ARM cores and FPGAs. These devices can perform aggregation, filtering, key-value store operations, and even machine learning inference on packets in flight, without delivering the packets to a host CPU. The result is a blurring of the boundary between "network" and "computer" that challenges the traditional layered architecture of the internet.\n\nThe deeper systems point is that in-network computing is a form of edge computing at the micro-scale: it pushes computation to the data not because the edge is "closer to the user" but because the edge is "closer to the data's path." The switch is not a passive router; it is an active compute node. The network is not a pipe; it is a distributed processor.\n\nIn-network computing is not an optimization. It is a recognition that the network has always been a computer — we just pretended otherwise to keep the abstractions clean.\n\n \n\nFuture directions include Co-Packaged Optics, the integration of optical transceivers with switch chips to eliminate the bandwidth bottleneck between switch and network, and Disaggregated Memory, the architectural shift that separates memory from compute nodes and requires in-network protocols to maintain cache coherence.