GPU Acceleration for 5G Signal Processing and Machine Learning

As the rollout of 5G progresses and research for 6G begins, the key themes of softwarization, virtualization, open systems and artificial intelligence form foundational principles for communication systems of the future. The application of AI/MLto wireless communications an extremely active research area with many 10's to 100's of papers published weekly reporting new results on the application of AI/ML to the physical layer (L1), MAC layer (L2) and at the network optimization level. To realize the Industry's vision of an AI/ML powered wireless future, a full stack solution supporting a software defined radio (SDR) approach for the vRAN, together with optimized silicon for AI, coupled with application development frameworks for AI/ML development is essential. NVIDIA GPU technology and associated CUDA programming model, together with arich suite of AI/ML SDKs (Software Development Kits) provides these capabilities. In this talk we present The Aerialsoftware-defined GPU-based cloud native 5G NR RAN platform. Aerial implements not only 5G NR the baseband signal processing, but using GPU virtualization supports additional concurrently operating workloads, such as AI/ML inference, training and data analytics on this one hyper converged system. We provide an overview of the L1 signal processing pipeline and describe efficient mechanisms for data movement between the GPU and NIC-based fronthaul interface using a GPU-enabled Data Plane Development Kit (DPDK). A brief survey of some of the promising deep learning approachesfor L1 and L2 enhancements is presented.