one6G at GLOBECOM 2025

At GLOBECOM 2025, held this year in Taipei, Prof. Iera – from CNIT (Italy) – presented the latest one6G paper entitled “GPU-Accelerated In-Network Computing for Split-AI in 6G: A Trade-Off Between Inference Time and Energy Consumption.”

The paper investigates In-Network Computing assisted Split-AI (INC-assisted Split-AI). The concept assumes distributing AI inference across multiple entities, including 6G user-plane entities with INC capabilities. The aim is to reduce AI inference latency, offload computational load from UE, or minimize network traffic. The study specifically focuses on the potential benefits of hardware acceleration for INC-assisted Split-AI in terms of inference latency.

Testbed experiments were conducted to assess (i) the achievable reduction in AI inference compute latency and (ii) the associated cost in terms of increased energy consumption. The results show that GPU acceleration can reduce Neural Network computation latency by up to 90%, while resulting in a comparatively moderate energy-consumption increase of no more than 22%.

The paper also evaluates the impact of GPU deployment location on the trade-off between inference time and energy consumption. The findings indicate that GPU placement can significantly influence this trade-off and should therefore be carefully considered by network operators.

Another scientific achievement from one6G WI207 “In Network Computing and Intelligent User Plane” from a joint effort by CNIT, Docomo, Huawei, ITAV and ITI.