AI-Native 6G: Empowering Intelligent RAN with Accelerated Compute

AI and communication is a key usage scenario envisaged for 6G. In contrast to 5G, where AI adoption has come only as an afterthought, 6G aims to embrace AI right from the beginning. AI-native RAN shifts the trend from fixed, one-size-fits-all algorithms to continuously adaptive, data-driven intelligence embedded throughout the RAN stack for site-specific optimizations. However, embedding AI for RAN introduces stringent latency constraints and computational challenges. In this article, we demonstrate how software-defined, accelerated compute platforms provide the foundation of AI-native 6G, focusing on AI-for-RAN which refers to using AI to enhance RAN performance. 

AI-native RAN must support real-time inference and must be adaptive to site-specific conditions. The comparison in Table I shows that while central processing unit (CPU), applicationspecific integrated circuit (ASIC), and field-programmable gate array (FPGA) each offer distinct advantages, none alone can unite rapid programmability, elastic compute scaling, lowlatency execution, and adaptability to new workloads or changing requirements – all together in diverse 6G scenarios. Software-defined, accelerated compute platforms with GPUs are able to bridge this gap.