Оптимізація мереж 5G засобами штучного інтелекту

Abstract

This work investigates modern methods for enhancing the performance of artificial intelligence algorithms executed directly on mobile and embedded devices. Specialized optimization approaches are considered, including layer-wise profiling, efficient execution of convolution operations using General Matrix Multiplication (GEMM), operator fusion to reduce memory accesses, low-precision computations (FP16, INT8), and memory usage optimization through liveness analysis and shared memory pools. The advantages of integrating modern deep learning architectures, such as Transformers and generative models (GANs), for network parameter prediction and training data synthesis are explored. Conclusions are drawn regarding improved inference speed, reduced memory footprint and energy consumption, and enhanced efficiency in utilizing GPU and AI accelerator resources.