In-Network Computing and Split-AI in 6G: Enablers and Proof-of-Concept Studies
Mattia Giovanni Spina, D. V. Soto Lebron, Susanna Schwarzmann, Riccardo Guerzoni, Riccardo Trivisonno, Antonio Iera, Floriano De Rango, George T. Karetsos, Thomas Zinner, Daniel Nunes Corujo · 2024
Split-AI proposes a paradigm shift in which Neural Networks (NN)-relevant tasks are distributed across multiple networking entities. The capabilities envisaged for 6G mobile networking such as In-Network Computing (INC) can facilitate a smooth deployment of this technique. 6G's Access Nodes (ANs) and User Plane Functions (UPFs) will be able to execute parts of an NN while performing their traditional functions (e.g. transmitting packets). Recent studies show that INC-assisted SplitAI has the potential to enhance Key Performance Indicators (KPIs) like inference time, network traffic load, and UE energy consumption. The incorporation of Split-AI in 6G networks needs to fulfill significant design and performance requirements such as optimal partitioning strategies for distributing NN parts across network elements. Additionally, the 6G User Plane must enable the dynamic and flexible allocation and deployment of these split NN parts. In this work we: i) propose possible enhancements for the Control Plane (CP) and UP, to make possible the integration of INC-assisted Split-AI in 6G networks; ii) present the design of a Proof-Of-Concept (PoC) simulation model aimed at demonstrating the feasibility of the technique with a comprehensive study on the impact that different NN partitioning options have on the network utilization, inference time, and the resources of both UE and network entities involved.