Benchmarking cross‑platform AI: Web Assembly, ONNX Runtime and TVM for Real‑Time Web, Mobile, and IoT Deployment

Chinnaraju, Aravind (2025) Benchmarking cross‑platform AI: Web Assembly, ONNX Runtime and TVM for Real‑Time Web, Mobile, and IoT Deployment. World Journal of Advanced Research and Reviews, 26 (2). pp. 1937-1963. ISSN 2581-9615

Cross‑platform deployment of machine‑learning inference now spans browser tabs, handheld applications, and resource‑constrained sensors, yet the performance landscape remains fragmented by heterogeneous runtimes. This study conducts the first holistic benchmark that positions WebAssembly, ONNX Runtime, and Apache TVM side‑by‑side under a unified test harness across Web, mobile, and IoT devices. A theoretical foundation distinguishes compilation from interpretation, ahead‑of‑time from just‑in‑time pipelines, and outlines how hardware‑abstraction layers mediate latency, throughput, memory, and energy trade‑offs. Empirical evaluations draw on a curated model zoo and cold‑start vs. steady‑state runs to expose four‑dimensional performance frontiers. Results show that TVM’s auto‑tuned kernels deliver up to a 42 % latency reduction on ARM microcontrollers, whereas WebAssembly narrows browser‑native overheads to within 1.4× of device‑bound baselines when SIMD extensions are available. ONNX Runtime provides the broadest portability, though execution‑provider selection must be coupled with quantization to remain within sub‑100 ms response budgets on mid‑tier smartphones. Integrating telemetry pipelines through OpenTelemetry and Delta Lake permits real‑time drift detection, AIOps‑driven auto‑rollback, and carbon‑aware scheduling that lowers energy use by 18 % without SLA violations. Security analysis contrasts browser sandboxes with enclave‑based protection for mobile and IoT, while risk‑management blueprints extend chaos‑engineering to runtime drift and compatibility faults. Case studies spanning a browser‑side image classifier, a mobile augmented‑reality pose estimator, and an IoT anomaly detector validate the decision matrix that maps workload characteristics to optimal runtime choice. The findings synthesise technical insights into actionable deployment playbooks, offering researchers and practitioners a reproducible framework for balancing performance, sustainability, and resilience in real‑time edge AI.