Published by Sean Carr
Asset Decode Was 83% of Boot. Parallelizing Textures Bought 5.6%.
A boot profile named the dominant phase; the first fix found the bulk of it was model decoding, not textures.
At a Glance
Asset decode and register
83.2%
Asset decode and register: 2.333 s = 83.2% of a 2.805 s boot.
First fix, texture decode only
-5.6%
Planned target: decode phase under 0.5 s. First version (texture decode only, byte-identical): -5.6%. Model loaders: about 66% of the decode phase, not yet parallelized.
Headless twin, asset decode
99.1%
Headless twin: asset decode 2.257 s = 99.1% of its boot. The windowed and headless figures agree within about 3%, pointing to CPU decode rather than device interaction.
GPU backend init
11.4%
GPU backend init: 0.320 s (11.4%). Audio device open: 0.141 s (5.0%).
| Telemetry Metric | Measured Value | Context & Observability |
|---|---|---|
| Asset decode and register | 83.2% | Asset decode and register: 2.333 s = 83.2% of a 2.805 s boot. |
| First fix, texture decode only | -5.6% | Planned target: decode phase under 0.5 s. First version (texture decode only, byte-identical): -5.6%. Model loaders: about 66% of the decode phase, not yet parallelized. |
| Headless twin, asset decode | 99.1% | Headless twin: asset decode 2.257 s = 99.1% of its boot. The windowed and headless figures agree within about 3%, pointing to CPU decode rather than device interaction. |
| GPU backend init | 11.4% | GPU backend init: 0.320 s (11.4%). Audio device open: 0.141 s (5.0%). |
| Independent decode bundles | 35-40 | About 35-40 independent decode bundles ran on one thread. |
Published by Sean Carr
View & print executive brief version →Executive Summary & Core Finding
A phase-stamped boot trace of the game found asset decode and registration took 2.333 s of a 2.805 s boot (83.2%). We planned a parallel decode with strictly serial registration to keep handle numbering and logs byte-identical. The first version warmed texture decode only and cut that phase by 5.6%, because model loaders turned out to be about 66% of it.
Method
Data collected Aug 1, 2026; written up October 2026.
August 1, 2026, canonical windowed Vulkan build (/O2, /fp:precise), warm file cache, one clean trace with phase stamps and under 5% unaccounted time.
A headless software-rendered twin was traced as a cross-check.
The follow-on design decoded assets on a boot-local thread pool while keeping registration on the boot thread in the original order, gated by a serial-parity A/B and golden-capture identity checks.
What We Measured
Phase-stamped boot trace and the first parallel-decode fix
Asset decode and register
83.2%
Asset decode and register: 2.333 s = 83.2% of a 2.805 s boot.
First fix, texture decode only
-5.6%
Planned target: decode phase under 0.5 s. First version (texture decode only, byte-identical): -5.6%. Model loaders: about 66% of the decode phase, not yet parallelized.
Headless twin, asset decode
99.1%
Headless twin: asset decode 2.257 s = 99.1% of its boot. The windowed and headless figures agree within about 3%, pointing to CPU decode rather than device interaction.
GPU backend init
11.4%
GPU backend init: 0.320 s (11.4%). Audio device open: 0.141 s (5.0%).
Independent decode bundles
35-40
About 35-40 independent decode bundles ran on one thread.
What We Think It Suggests
- Measuring the boot by phase before touching it identified one phase as the clear target. Measuring inside that phase would have shown earlier that model decoding, not texture decoding, was the larger share.
- Parallel decode with serial, order-preserving registration looks like a workable shape when handle numbers and logs are part of observable identity, but this run only validated it for textures.
Limits
- Single trace on one development machine with a warm cache; cold-cache and low-end hardware boots were not measured.
- The after-fix boot table was flagged in the source as not admissible (the machine was not quiet and another default had changed), so a clean before/after boot comparison is still owed.
- The plan header also cites an earlier audit figure of 88% of boot from a different measurement; this article uses the 83.2% traced figure.
Detailed Analysis & Field Observations
A phase-stamped boot trace of the game found asset decode and registration took 2.333 s of a 2.805 s boot (83.2%). We planned a parallel decode with strictly serial registration to keep handle numbering and logs byte-identical. The first version warmed texture decode only and cut that phase by 5.6%, because model loaders turned out to be about 66% of it.
Data collected Aug 1, 2026; written up October 2026.