Dynamic reconfigurable RF system for MIMO 5G applications
Lead Research Organisation:
University of Edinburgh
Department Name: Sch of Engineering
Abstract
The aim of the project would be to develop and optimise the latest algorithms targeting 5G communication on the latest reconfigurable platform designed by Alpha Data, based on the Xilinx Zynq RFSoC device that integrates ARM CPUs, FPGA Fabric and RF-ADC/RF-DAC into a single IC. Specific algorithms that will be considered will be those targeting multiple antenna systems requiring sophisticated beamforming techniques on-the-fly, that can be dynamically changed through reconfiguration of the FPGA DSP Fabric and RF interface. The algorithms will be optimised in terms of area, power, and performance.
The research will be done in cooperation with Alpha Data, a leading supplier of COTS FPGA acceleration and interface cards for the Embedded and Data Centre markets. Based in Edinburgh since 1993, and Denver since 2007, the company specializes in providing early to market reconfigurable computing modules that can be used to develop solutions and then deploy them on the same hardware. Customers range from hyper-scale data centres and major communications, aerospace and defence giants to machine learning start-ups and academic and scientific institutions.
The research will be done in cooperation with Alpha Data, a leading supplier of COTS FPGA acceleration and interface cards for the Embedded and Data Centre markets. Based in Edinburgh since 1993, and Denver since 2007, the company specializes in providing early to market reconfigurable computing modules that can be used to develop solutions and then deploy them on the same hardware. Customers range from hyper-scale data centres and major communications, aerospace and defence giants to machine learning start-ups and academic and scientific institutions.
People |
ORCID iD |
| T Arslan (Primary Supervisor) | |
| Stefan Brennsteiner (Student) |
| Description | Since roughly 2017, machine learning methods have been proposed to aid the physical layer signal processing in communication systems to increase performance and reliability. These proposals largely rely on simulation and the computational complexity of the proposed algorithms is often prohibitively large. Our recent, work shows that by applying neural network compression techniques the complexity of model-driven designs can be significantly reduced. Combined with high-performance custom circuits on field-programmable gate arrays, we achieve real-time processing of significant bandwidths in a Long Term Evolution (LTE) /5G-like communication system for the first time. We believe that these results are an important step towards deploying a full machine learning-based LTE/5G physical layer processor in the future, unlocking the performance and reliability gains shown by simulation in the literature. |
| Exploitation Route | We hope that our work will serve as a basis for designing a full, more complete machine-learning-based physical layer processor. Also, we hope that it will help to recognize the importance of the computational complexity of these algorithms, especially for data-driven approaches. Only deployable algorithms will have an impact outside of academia. |
| Sectors | Digital/Communication/Information Technologies (including Software) |