Your GPU Is Up. Is Your AI? The Inference Resilience Profile
Abstract
Healthy GPUs and available infrastructure do not, by themselves, demonstrate that an AI service will remain usable during failure. This whitepaper proposes an Inference Resilience Profile and a companion benchmark, ResilientInferenceBench, for evaluating AI service continuity across Kubernetes and NVIDIA GPU failure domains. The proposed profile defines workload characteristics, failure domains, state continuity policies, isolation requirements, reserve capacity, service objectives, and evidence requirements. The benchmark design covers controlled failure, degradation, contention, and maintenance scenarios. It evaluates request survival, output validity, stream integrity, retained throughput, latency, recovery behavior, fault propagation, tenant isolation, cost, and energy consumption. The proposal connects existing infrastructure, observability, and benchmarking evidence to explicit service continuity requirements. It is intended for enterprise architects, platform engineers, AI SRE teams, infrastructure operators, security teams, and researchers working on resilient AI systems. Version 0.1.3 is a public draft of a technical proposal. It has not undergone formal external peer review or empirical platform validation. Numerical examples are synthetic illustrations, not measured benchmark results. Companion software, schemas, and executable test tooling are planned and are not included in this deposit. Future validation is intended to include at least two materially different platforms, reproducible methods, and published evidence. AI assistance disclosure: OpenAI Codex and OpenAI image generation tools assisted with manuscript editing, technical consistency checks, document formatting, and the generation and revision of selected figures. The author is responsible for the published content.
