CREST-LLM: Carbon–Reliability–Performance Evaluation Framework for Sustainable LLM Systems

Large Language Model (LLM) inference systems are increasingly deployed at scale, yet existing hardware evaluation methodologies either prioritise performance or consider sustainability dimensions in isolation, without accounting for LLM-specific inference behaviour and reliability-driven lifetime effects. This paper presents CREST-LLM, an integrated evaluation framework for LLM inference that jointly models performance, reliability, and lifecycle carbon at design time. It integrates (i) a phase-aware inference energy model, (ii) a lifecycle carbon model capturing embodied and operational emissions, and (iii) an electromigration-based ageing model to quantify long-term reliability. Through simulation-driven exploration of heterogeneous architectures, we show that latency-optimized designs can incur up to 2× higher lifetime carbon and that ageing can reduce system lifetime by over 30%. We further show that selective redundancy extends lifetime while remaining near carbon break-even, providing actionable guidance for sustainable LLM system design.