Leverage Mobile GPUs for Flexible High-Speed Wireless Communication

Due to the fast emergence of new wireless technologies, traditional hardware accelerator based wireless baseband processors can no longer meet the requirement of a short time to market. On the other hand, general-purpose processors support good general programming models, which can greatly reduce the development cycle and effort for new mobile devices to support the most advanced wireless technologies. Among all the general-purpose platforms, graphics processing units (GPUs) are designed to achieve high throughputs and high energy efficiency by exploring data-level parallelism and thread-level parallelism. In addition, general-purpose programming support has been added to mobile GPUs recently. These make a mobile GPU a promising candidate for the flexible high-speed wireless communication in a user device. In this work, we evaluate using a mobile GPU for the LTE baseband downlink receiver in a mobile device. We first develop highly parallel CUDA implementations of key signal processing kernels of an LTE downlink receiver. Then we evaluate achieved throughputs and latencies when running an LTE downlink receiver on an NVIDIA Jetson Tegra K1 board. Our results show that by deploying packet batching, we can meet LTE specification peak throughputs for all but one configuration for the physical layer kernels. We also profile the kernels, and point out that the demodulation kernel is the hotspot because of its high computation. Finally, we evaluate the energy efficiency of the mobile GPU under different frequencies, and show that a high frequency achieves better energy efficiency. Compared with other baseband processors, a mobile GPU still has relatively high energy consumption. We provide several implications for the future mobile GPU design to reduce the energy consumption, including having ISA supports for the Turbo decoder and reducing the energy waste from unused registers.