Published by Sean Carr
A Stale Safety Comment Parked a 10 ms Win for Three Days
The guard against GPU skinning was correct when written; nobody updated it after the two bugs it guarded against were fixed.
At a Glance
Frame time, scalar vs GPU skinning
-67.4%
Frame time: scalar 15.520 ms, GPU 5.064 ms (-10.456 ms, -67.4%); no-skinned-body floor 3.587 ms; SIMD path 25.856 ms.
Scalar vs GPU image diff
1 differing pixel
Scalar vs scalar: 0 differing pixels of 3,686,400. Scalar vs GPU: 1 differing pixel, max channel delta 1.
Per-body-pair skin cost
11.933 ms to 1.477 ms
Per-body-pair skin cost: 11.933 ms to 1.477 ms.
Wrong records of this default
four were wrong
Records of the default for this one switch: five; four were wrong.
| Telemetry Metric | Measured Value | Context & Observability |
|---|---|---|
| Frame time, scalar vs GPU skinning | -67.4% | Frame time: scalar 15.520 ms, GPU 5.064 ms (-10.456 ms, -67.4%); no-skinned-body floor 3.587 ms; SIMD path 25.856 ms. |
| Scalar vs GPU image diff | 1 differing pixel | Scalar vs scalar: 0 differing pixels of 3,686,400. Scalar vs GPU: 1 differing pixel, max channel delta 1. |
| Per-body-pair skin cost | 11.933 ms to 1.477 ms | Per-body-pair skin cost: 11.933 ms to 1.477 ms. |
| Wrong records of this default | four were wrong | Records of the default for this one switch: five; four were wrong. |
| Body drawn | 38,240 px | Body drawn: 38,240 px on both arms with identical bounding box. |
Published by Sean Carr
View & print executive brief version →Executive Summary & Core Finding
A code comment barred switching character skinning to the GPU until an A/B showed the body rendered correctly. It was written when that was true, then outlived two fixes by three days. Re-running the A/B it demanded showed 1 differing pixel in 3,686,400 and a frame-time drop from 15.520 ms to 5.064 ms.
Method
Data collected Aug 7, 2026; written up October 2026.
August 7, 2026, live windowed Vulkan lane, one binary with the skinning path selected by an environment switch and wind phase pinned so animation could not masquerade as an arm difference.
Image comparison was against a scalar-vs-scalar noise floor, with a separate signal confirming the GPU path actually engaged.
Performance used 4 paired 600-frame legs, each verified to have presented all 600 frames.
Hardware is not recorded in the commit.
What We Measured
Re-running the A/B a stale comment demanded
Frame time, scalar vs GPU skinning
-67.4%
Frame time: scalar 15.520 ms, GPU 5.064 ms (-10.456 ms, -67.4%); no-skinned-body floor 3.587 ms; SIMD path 25.856 ms.
Scalar vs GPU image diff
1 differing pixel
Scalar vs scalar: 0 differing pixels of 3,686,400. Scalar vs GPU: 1 differing pixel, max channel delta 1.
Per-body-pair skin cost
11.933 ms to 1.477 ms
Per-body-pair skin cost: 11.933 ms to 1.477 ms.
Wrong records of this default
four were wrong
Records of the default for this one switch: five; four were wrong.
Body drawn
38,240 px
Body drawn: 38,240 px on both arms with identical bounding box.
What We Think It Suggests
- A blocking comment is a claim about the world, and it needs the same expiry discipline as any other claim. This one encoded a valid precondition but not a way to notice when the precondition stopped being true.
- Confirming that the new path actually ran mattered as much as the pixel diff: a fail-soft fallback would have measured the old path twice and reported a perfect match.
Limits
- One scene, one character setup, one machine; the 10 ms figure depends on how much skinned geometry is on screen.
- Four paired legs is a small sample, though the effect is large relative to the differences between legs we observed.
- Hardware is not recorded in the source.
Detailed Analysis & Field Observations
A code comment barred switching character skinning to the GPU until an A/B showed the body rendered correctly. It was written when that was true, then outlived two fixes by three days. Re-running the A/B it demanded showed 1 differing pixel in 3,686,400 and a frame-time drop from 15.520 ms to 5.064 ms.
Data collected Aug 7, 2026; written up October 2026.