Hardware footprint benchmarks for an offline maps/routing/geocoding stack

Context: I posted a Show HN a while back about an offline maps stack (vector tiles + routing + geocoding) for edge/on-prem/air-gapped deployments, and one common question was about hardware requirements.

Original thread for background: https://news.ycombinator.com/item?id=46478061

This post is a write-up of measurements across multiple regions on a consistent baseline (EC2 c6g.xlarge / Graviton2), including methodology + raw numbers (disk footprint, p95 RAM, and p95 endpoint latencies).

If you’ve deployed mapping/search/routing into constrained or secured environments, what would you want benchmarked to make a real go/no-go decision?

Examples: Raspberry Pi/NUC class hardware, cold-start, concurrency, update workflow, custom overlays/POIs, address quality, terrain/elevation (hillshade/contours/LOS).

If you share rough "region size + target hardware + usage pattern," I’m happy to add more rows.