Six Blender shards rendered the 3D library thumbnails 8 percent faster and 6 percent more energy per picture than the one-process pilot, 13.2 against 12.2 pictures per second and 4.99 against 4.70 joules per picture.
The 3D library thumbnail run used six headless Blender shards. Our plan asked for one resident process. Against the one-process pilot, the shards made pictures 8 percent faster, at 13.2 against 12.2 per second. They also used 6 percent more energy per picture, at 4.99 against 4.70 joules. The run finished 3,884 pictures in 294 seconds and about 19 kilojoules.
What the library held
The library had 3,972 assets in the end: 2,853 3D files and 1,119 Blender files. The plan had counted 1,130 Blender files. The run found 1,119. We do not know why the plan's count was higher, and the record does not say.
The plan asked for one resident headless Blender process, which imports each file, frames it, renders a 200 pixel picture, deletes the scene and moves on. The run split the work across six processes instead.
The run and the pilot
The timed run covered 3,884 pictures in 293.9 seconds. The tool reports package energy of 19,392 joules, an average of 66 watts, which is 4.99 joules per picture. The one-process pilot read 4.70 joules per picture and 12.2 pictures per second.
The 8 percent speed gain and the 6 percent energy cost are both small. They are also the only comparison we have. The sources give no spread for either figure, so we cannot say how much of the gap is noise.
What the plan asked for
The plan split the library into six kinds and gave each kind a method. Videos, audio, SVG and textures ran on the CPU, each with a fixed tool. Blender files used the preview stored inside the file when one existed. Files without one were to be counted and reported before anyone chose another method. The 2,853 3D files ran in one resident headless Blender process: import, frame to the bounding box, render with Workbench and texture colour at 200 pixels, delete the scene, and move to the next. The plan named the GPU for that step, because Workbench opens the device.
Each finished thumbnail is stored under the SHA-256 hash of its source file, at 200 by 200 pixels in RGB. The hash is the cache key, so a rerun can skip the files it already has. The timed run did not follow the one-process step, and the rest of this article is about that gap.
What failed
Four Blender files failed with enum "PNG" not found in ('FFMPEG'). Each had been saved for video output, and the glTF import read from it. Setting the image format to IMAGE fixed all four on retry.
Thirty files failed with "no visible mesh-like object". That is a clear reason. A further 53 failed with an empty reason string. The record gives no explanation for those 53, so we report them as unexplained.
The final tally was 3,888 thumbnails with 84 assets left without one, which sums to the 3,972 assets. The 3,888 includes the four fixed on retry. The timed run's 3,884 does not.
What the speed comes from
The record compares the run with a one-process pilot. It does not list every setting the two shared, so we cannot rule out other differences. The shard count is our explanation for both the gain and the cost. We have not run a one-shard arm on the same list, so it is not yet a measured one.
What we still do not know
Why the 53 files failed. Whether the 84 assets left without a thumbnail are the same files as the 88 without one at the first pass (thumbs.md line 3); the record does not list them. Whether the pilot ran under the same machine load as the timed run. What the energy reading covers beyond the package. Whether a plan that follows the brief to one process would be slower or cheaper. We would rather run that than guess.
How to prove this wrong
Run the same asset list twice on the same machine within a short window: once with one shard, once with six. Keep the GPU state, the Blender version and the output size fixed. Report both pictures per second and joules per picture with the spread from a repeat. If the one-shard run is within that spread of six, the shard count is not the cause, and the gain has a different explanation.
Provenance
/home/mario/Projects/AssetsHQ/.work/thumbs/thumbs.md, line 3 (3,972 3D and Blender assets at the first pass), line 5 (the run: six shards, 293.9 s wall, 19,392 J package energy, 4.99 J per picture, 13.2 pictures per second; the one-process pilot at 4.70 J and 12.2 per second), lines 19 and 20 (53 failures with an empty reason; 30 with no visible mesh-like object), lines 59 and 60 (final tally of 3,888 thumbnails and 84 without; four retries with the image format set to IMAGE).
/home/mario/Projects/AssetsHQ/docs/lanes/library-thumbs.md, lines 11 to 18 (the plan: 1,130 Blender files, 2,853 3D files, one resident headless Blender process).
Four steps at 768 pixels cut a Z-Image picture from 7.1 seconds to 1.7. The sampler is already near the card's compute limit, so the gain came from steps and size, and the texture loss is the price.
Re-rolling two soft plates at a larger size with more steps and new words gave the approved set. Our sources do not separate those changes, so the article says what changed and what the next test has to be.
The prediction assumed the model and encoder swaps stayed in every arm. On cache-hit shots they did not, and the warm walls landed near 24 seconds. Real time is still about 8 to 10 times too slow, so the cache is not a path to the goal.
Commenti
Ancora nessun commento.
Accedi per commentare.