Skip to main content

Synthetic timing evidence

Back to Grid. Recorded 2026-11-12T23:13:16.425Z. This is a non-geographic compute probe, not a validated Earth production pipeline, visual quality assessment or flight performance benchmark.

Measured stages​

Five sequential runs per profile, no discarded warm-up. Each range is observed minimum–maximum seconds, not a confidence bound. Total includes preparation, export and validation; stage extrema occur on different runs and must not be added.

TrianglesPreparationExport/writeValidation/readback
32,7680.012–0.0270.005–0.0090.012–0.025
131,0720.046–0.0580.016–0.0220.037–0.057
524,2880.174–0.1780.052–0.0790.140–0.168
VerticesTextureGLB bytesTotal rangeMedian
16,641512² RGB1,715,3040.03–0.06 s0.04 s
66,0491024² RGB6,839,9960.11–0.14 s0.12 s
263,1692048² RGB27,321,3480.37–0.42 s0.39 s

Preparation creates a planar 72,000-unit square sinusoidal height mesh, analytic normals, UVs and seeded RGB noise. PNG level 6 encoding is included. This deliberately does not sample real Earth, project S2 corners or test curvature. Noise texture and mesh settings are explicit test workloads, not selected terrain quality presets. Export uses glTF Transform NodeIO; geometry is uncompressed, textures are PNG, no Draco/Meshopt/KTX2. Validation uses Khronos glTF Validator plus a full NodeIO readback checking vertex/index counts and texture width. All 15 files had zero format errors and zero warnings, with identical hashes within each profile. Files were deleted immediately after measurement; no model is published. No rendered visual-quality acceptance is claimed.

Hardware and limits​

Intel(R) Core(TM) i9-14900KS, linux x64, Node v26.12.1; container CPU quota 800000 100000 (8 CPU equivalents), memory limit 34359738368 bytes (32 GiB). One serial worker, Sharp concurrency 1. glTF Transform 4.5.0, Sharp 0.35.5, Khronos validator 2.0.0-dev.3.10. This is a shared host, not a controlled dedicated machine; five repetitions cannot establish production tails.

Timing begins after imports/process startup. Temporary filesystem writes do not call fsync and may use OS cache; this is not a durable disk or object-store throughput test. Memory capacity is a limit, not measured peak consumption. No parallel scaling is inferred.

Excluded: downloading DEM/imagery, mosaicking/reprojection, real source decoding, spherical placement, terrain seams, collision/content QA, production geometry/texture compression, upload, retries and human review. Real source-to-geographic-GLB time is unknown. These timings cannot justify a continent schedule or an ordinary-PC flight-quality claim.

Reproduce and inspect​

With the repository's frozen pnpm dependencies installed, run from its root:

node scripts/benchmark-synthetic-glb.mjs > ../synthetic-repeat.json
node scripts/earth-grid-pages.mjs --check
pnpm test

Use a writable scratch directory via TMPDIR; the script creates a unique temporary subdirectory and removes it in a finally block. It runs exactly 15 files, maximum 524,288 triangles / 2048² texture, one temporary GLB at a time. It never fetches geographic data or writes production assets. Wall times vary; compare geometry, file hashes and validation results, not identical durations. The committed raw evidence is research/earth-grid/synthetic-benchmark.json; the executable method is scripts/benchmark-synthetic-glb.mjs. Generated reader pages consume this evidence directly, and tests recalculate min/median/max from every raw run.