RadioML Meets FINN: Enabling Future RF Applications With FPGA Streaming Architectures

Abstract

Deep neural networks (DNNs) are penetrating into a broad spectrum of applications and replacing manual algorithmic implementations, including the radio frequency communications domain with classical signal processing algorithms. However, the high throughput (gigasamples per second) and low latency requirements of this application domain pose a significant hurdle for adopting computationally demanding DNNs. In this article, we explore highly specialized DNN inference accelerator approaches on field-programmable gate arrays (FPGAs) for RadioML modulation classification. Using an automated end-toend flow for the generation of the FPGA solution, we can easily explore a spectrum of solutions that optimize for different design targets, including accuracy, power efficiency, resources, throughput, and latency. By leveraging reduced precision arithmetic and customized streaming dataflow, we demonstrate a solution that meets the application requirements and outperforms alternative FPGA efforts by 3.5Â in terms of throughput. Against modern embedded graphics processing units (GPUs), we measure > 10Â higher throughput and > 100Â lower latency under comparable accuracy and power envelopes.

An Exploration of State-of-the-Art Automation Frameworks for FPGA-Based DNN Acceleration

“…They propose estimation models to show how the quantization affects the hardware cost and throughput. Jentzsch et al [11] compare the performance of FINN-based accelerators with GPUs on the RadioML modulation classification task. They extend FINN to support additional parallelism.…”

“…In this paper, we aim to answer the above questions by conducting an in-depth exploration of the performance and power efficiency characteristics of DNN accelerators generated by FINN and Vitis AI, two prominent representatives of the two automation framework groups. There have been some efforts to evaluate FINN and Vitis AI individually, with the comparison targets being the executions on CPUs or GPUs [11], [12], [13]. Investigating the gap between these two frameworks, however, is still an open question.…”