D2.1 Evaluation of GPU Implementation of ML-based Receiver

An important aspect of CENTRIC project is the practical deployment and evaluation of the developed algorithms on the target hardware platforms. In this report, we summarize our work on Task T2.2.1 related to real-time inference of ML-based components. We have implemented a real-time neural version of our network (NN)-based multiuser multiple-input multiple-output (MU-MIMO) receiver with 5G New Radio (5G NR) compatibility using TensorRT, high-performance deep learning inference library. We havemeasured the real-time latency for the target platform, which is a key performance indicator for practical 5G NR systems. Further, we have also refined our training strategies based on a multi-loss function that allows us to control the latency of the receiver for a given hardware target platform by balancing the computational complexity and the accuracy of the NN. This enables us to adjust and reconfigure the latency for new hardware platforms without re-training, which is a desirable feature for future-proof and flexible AI/ML solutions. We have evaluated the latency versus block error rate (BLER) performance trade-off for different NN architectures and configurations and concluded that the computational latency is a crucial system parameter that needs to be considered when comparing different AI/ML receiver algorithms.