EXPLORATORY INTERNAL WORKFLOW COMPARISON
4014 Labs and Blender on Modal
Three founder-run tests of the delay between an initial request and a delivered file.
2.1-million-polygon scene: 20-second 1080p video
- 4014 Labs
- 67.215 s
- Blender on Modal
- 279.474 s
Blender on Modal took 4.16× as long in this internal test.
Internal results using different rendering engines and different GPU hardware. Reproduction materials are not yet published.
Results
| Output | 4014 Labs | Blender on Modal |
|---|---|---|
| 3D-render workflow: 1080p video, 600 frames | 67.215 s | 279.474 s |
| Minimal-output latency probe: black 512 × 512 PNG | 6.639 s | 23.161 s |
| Minimal-output latency probe: black 256 × 256 one-frame MP4 | 4.360 s | 21.651 s |
Test setup
3D-render workflow
- The long render was 1080p, 600 frames, and 20 seconds long.
- The scene contained 2.1 million polygons.
- The 3D scene was modeled on CPU in Python.
- The tested 4014 Labs workflow used an NVIDIA P4. The tested Blender-on-Modal workflow used an NVIDIA L4.
- The Modal workflow ran Blender. The 4014 Labs workflow used our rendering engine.
- Timings include setup, rendering, encoding, and delivery. They are not raw GPU-kernel timings.
Minimal-output latency probes
- The PNG probe measured first contact to a black 512 × 512 image.
- The MP4 probe measured first contact to a black 256 × 256, one-frame video.
- These probes did not use the 2.1-million-polygon scene described above.
What the benchmark tests
This is a comparison of complete product workflows, not equivalent rendering engines or equivalent hardware. It measures the delay an agent encounters when it asks for a file and waits for the result.
In the founder-run Blender-on-Modal workflow, we estimated 120.47 seconds during which the reserved GPU was inactive. The derivation of that estimate is not yet published. Our workflow preprocesses on CPU before rendering on GPU.