Next-Gen AI-on-RAN: AI-Native, Interoperable, and GPU-Accelerated Testbed Towards 6G Open-RAN

The evolution of 5G and the emergence of 6G networks demand advanced, artificial intelligence-enabled (AIenabled), interoperable, and scalable solutions to overcome the ever-increasing complexity and performance requirements of nextgeneration wireless communication systems. To address these design challenges, we propose a Next Generation AI on Radio Access Network (Next-Gen AI-on-RAN) testbed an open-source platform that leverages NVIDIA Aerial RAN CoLab (ARC) and Aerial Software Development Kit (SDK) to enable virtualized Graphics Processing Unit (GPU) accelerated Layer 1 (L1) and Layer 2 (L2) processing, significantly enhancing computational efficiency and reducing latency. A novel aspect of this work is the co-location of training and inference workloads on the same NVIDIA GPU-centric RAN server, wherein the GPU is simultaneously used for real-time L1/L2 processing and AI model training for traffic prediction. Experimental analysis demonstrates that our design 32% reduced training time without sacrificing model accuracy which underscores the potential of GPU-accelerated AI-on-RAN for robust next-generation wireless networks.