Prototyping Real-Time Semantic Communications With SDRs and Edge GPUs

Semantic communication represents a paradigm shift by prioritizing task-relevant information over bit-level accuracy. Despite recent progress, most studies are predominantly restricted to simulations. Even works involving practical deployment often lack detailed hardware-software integration, flexible architectures for diverse semantic algorithms, and validation under low signal-to-noise ratio conditions. To address these gaps, this letter presents a flexible and scalable prototype that integrates software-defined radios (SDRs) with edge graphics processing units (GPUs) to enable real-time semantic processing and over-the-air transmission. The system adopts a modular architecture that supports customizable frame structures, multimodal data, and rapid deployment for evaluating diverse semantic algorithms. It further enables the integration of large-scale foundation models and experimentation in complex wireless environments. Hardware and software co-design are detailed, and the prototype is validated through three representative case studies: semantics-driven frame synchronization, image semantic transmission, and foundation model-aided semantic communication. In addition, inference latency and bandwidth overhead analyzes provide quantitative guidance for algorithm design. Experimental results demonstrate that the proposed prototype operates reliably in complex wireless environments and serves as a practical testbed for real-world semantic communication research and development.