EcoAccel-ITAD: A Telemetry-Driven Diagnostic and Embodied Carbon Accounting Framework for Second-Life Heterogeneous AI Accelerators

The explosive expansion of enterprise AI infrastructure has shortened replacement cycles for high-density AI accelerators (GPUs, NPUs, and TPUs), creating crit ical e-waste challenges and poorly quantified Scope 3 carbon emissions. Unlike general-purpose CPUs, AI accelerators experience severe thermal cycling, high bandwidth memory (HBM/GDDR) degradation, and power-delivery degradation under continuous matrix multiplication loads. This paper presents EcoAccel ITAD, a telemetry-driven, non-destructive diagnostic and carbon allocation framework tailored for second-life heterogeneous AI hardware at the point of decommis sion. EcoAccel-ITAD interfaces directly with vendor management APIs (Nvidia NVML, AMD ROCm SMI, Intel OneAPI Level Zero) to extract low-level degrada tion metrics, including uncorrectable ECC error counts, memory controller clock throttling, and silicon aging proxies. These parameters formulate a composite Ac celerator Health Index (Saccel), which maps directly to an Arrhenius-weighted remaining-service-life model for quantifying avoided embodied carbon (Csaved). Evalu ated across 50 executions on an enterprise tensor pro cessing testbed, the system completed full diagnostic assessment, sanitization verification, and cryptographic certificate generation in 2.189 s (σ = 0.015 s) with peak memory overhead of 18.4 MB and zero persistent writes to host storage.