Pillar tour II: reliability, performance, cost, sustainability
Lesson 4 of 5 in The Generative AI Lens: AWS Well-Architected for GenAI.
The remaining four pillars are shorter but each carries a distinctly GenAI twist that the base framework never had to think about. Read them as four answers to one question: what changes when the workload’s core component is a large, expensive, rate-limited, probabilistic model?
Reliability
The scarce resource is model throughput. GENREL01 opens with a question no classic workload needed: how do you determine throughput quotas for foundation models? Managed and serverless endpoints can only service so many inference requests at a time, so the best practice is to scale and balance model throughput as a function of utilisation — watch the quota headroom the way you once watched CPU.
The rest of the pillar handles the consequences: reliable network paths between components, graceful failure when a model call errors or throttles (retries, fallbacks, honest degraded modes — a chatbot that says try again shortly beats one that hallucinates through a timeout), prompt management with version control so a prompt rollback is as routine as a code rollback, and distributed availability across regions with verification that long-running distributed compute tasks actually completed.
Performance
Task performance joins latency and throughput. GENPERF01 asks how you capture and improve model performance in production — and its first best practice is to define a ground-truth dataset of prompts and responses, because a foundation model’s strength in one task (summarisation) says nothing about another (question answering). Benchmarks must be yours.
The pillar’s most operational insight is GENPERF04: the retrieval layer is a first-class performance component. A slow or badly tuned vector store drags the whole system, with cascading effects that are hard to attribute. Best practices: test vector embeddings for latency and relevance, and optimise vector dimensions for your use case — bigger embeddings are not automatically better, just automatically slower and costlier to search.
Cost
The prompt is a line item. Token-metered pricing makes cost a property of every request, so the lens treats prompt engineering as a cost discipline: GENCOST03’s best practices are to optimise prompt token length, control response length, use prompt caching so repeated context is not re-billed at full rate, and annotate user input for cost-aware filtering — screen cheap and early rather than expensive and late.
Upstream of that, GENCOST01 makes model selection a cost decision: right-size the model to the task, because a smaller model that clears your quality bar beats a frontier model that clears it identically at several times the price — your GENOPS01 evaluation harness is what makes that comparison honest. The pillar closes with workflow boundaries: agent loops and vector pipelines need explicit exit conditions, or a confused agent becomes an unbounded spend.
Sustainability
The footprint question starts before the architecture. The pillar’s first principle is bracing: ask whether generative AI is the right solution at all — a search engine answers a lookup with a fraction of the resources a generative model burns creating new text. That scoping question is a sustainability control.
Given a GenAI answer, the practices are efficiency-shaped: auto scaling and serverless architectures so compute exists only while used, efficient model customisation services rather than bespoke training runs, and energy-efficient models — smaller, distilled, or quantised models on efficient hardware. Notice the alignment: nearly every sustainability practice is also a cost practice, which is why these two pillars tend to pass and fail together in reviews.
Interactive sorting exercise: A reviewer hears each of these during a lens review. Sort each concern into the pillar whose questions address it.
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.