Semantic Communications and Edge Intelligence in 6G Wireless Networks: Task-Oriented Joint Source-Channel Coding and Latency Bounds

Abstract—As cellular architectures transition from 5G toward Sixth-Generation (6G) wireless networks, the classical Shannon-Nyquist paradigm of bit-level data transmission is rapidly approaching its theoretical physical channel capacity limits. Ultra-dense Internet of Things (IoT) ecosystems, autonomous vehicular swarms, and extended reality (XR) telepresence demand terabit-per-second (Tbps) data rates, sub-millisecond latencies, and continuous real-time edge artificial intelligence (AI) inference over severely bandwidth-constrained wireless channels. Classical communications optimize bit error rates (BER), transmitting massive streams of uncompressed binary representations regardless of the underlying computational task. Semantic Communications and Task-Oriented Edge Intelligence establish a transformative paradigm shift: shifting the transmission objective from error-free bit reconstruction to the faithful transmission of actionable semantic meaning and task-relevant latent features. By uniting Deep Joint Source-Channel Coding (Deep JSCC) with transformer-based feature extractors, semantic communication frameworks filter out redundant sensory data at the transmitter, communicating only task-critical representations to edge compute nodes. However, deploying semantic communications over stochastic wireless channels introduces acute foundational challenges: semantic noise induced by channel fading, joint training instability across non-differentiable wireless channels, and catastrophic performance degradation under out-of-distribution physical channel shifts. This paper presents an exhaustive, mathematically formal technical investigation and empirical benchmark of semantic communications for 6G edge intelligence. We formalize the rate-distortion-perception tradeoff under Weaver-Shannon semantic communication tiers, model continuous-time Deep JSCC over Rayleigh and Rician fading channels, and establish Lyapunov-guided end-to-end latency bounds for distributed edge inference. Furthermore, we construct an empirical software-defined radio (SDR) and GPU testbed evaluating multimodal perception and edge autonomous vehicle telemetry under severe signal-to-noise ratio (SNR) degradation. Our empirical findings demonstrate that Deep JSCC semantic transmission achieves a 78.4% reduction in required channel bandwidth while preserving 94.6% downstream edge classification accuracy under extreme channel noise (SNR = -6 dB), where classical JPEG/264 paired with LDPC coding completely collapses. Finally, we formulate open research directions in semantic knowledge base synchronization, neuro-symbolic semantic verification, and zero-trust semantic cybersecurity. Index Terms—Semantic Communications, 6G Wireless Networks, Edge Intelligence, Deep Joint Source-Channel Coding (Deep JSCC), Task-Oriented Communications, Shannon-Weaver Levels, Physical Layer AI.