cloth_shirt_4
Four transparent garment meshes from unisex_shirt.usd.
Cross-system benchmark results plus orbiting visual comparisons for transparent cloth, robots, and repeated mesh-heavy scenes. The visual comparison viewer updates frame-by-frame with a mouse-draggable timeline.
| System | GPU | CPU | RAM | OS | Driver | Framebuffer | Newton | Warp | Python |
|---|---|---|---|---|---|---|---|---|---|
| RTX 4090 desktop | NVIDIA GeForce RTX 4090, 24 GiB | Intel Core i9-12900K | 64 GiB | Windows 11 10.0.22635 | 595.79 | 1920 x 1080 | 1.4.0.dev0 | 1.14.0 | 3.12.3 |
| MacBook Pro M1 | Apple M1, 8-core integrated GPU | Apple M1, 8 cores (4 performance, 4 efficiency) | 8 GiB unified memory | macOS 26.2 (25C56), arm64 | OpenGL 4.1 Metal - 90.5 | 3840 x 2160 (1920 x 1080 request, Retina 2x) | 1.4.0.dev0 | 1.14.0 | 3.12.12 |
cloth_shirt_4 OIT speedup. Sorted and OIT are effectively tied on this view.
shape_cloud_240 OIT speedup over sorted blending.
mesh_spheres_160 OIT speedup over sorted blending.
bunny_meshes_24 OIT speedup over sorted blending.
cloth_shirt_4 OIT speedup. OIT takes 16.152 ms versus 13.541 ms sorted.
shape_cloud_240 OIT speedup. OIT takes 12.519 ms versus 9.801 ms sorted.
mesh_spheres_160 OIT speedup. OIT takes 12.313 ms versus 8.889 ms sorted.
bunny_meshes_24 OIT speedup. OIT takes 11.386 ms versus 8.190 ms sorted.
Sorted blending draws transparent objects back-to-front and uses conventional alpha blending. Weighted OIT accumulates transparent fragments into order-independent buffers and resolves the result approximately. The OIT path avoids object-level ordering sensitivity, while visual differences remain scene-dependent.
Reading the results: OIT speedup is
(sorted mean / OIT mean - 1) x 100, so negative values mean OIT is slower.
The M1 requested a 1920 x 1080 window, but the native macOS Retina context produced a
3840 x 2160 backing framebuffer. It therefore shaded four times as many pixels as the RTX run;
compare each system's sorted-versus-OIT behavior rather than treating the two systems as equal-resolution tests.
| System | Scene | Geometry | Sorted mean ms | OIT mean ms | OIT speedup | Mean abs RGB | RMS RGB | Max abs RGB |
|---|---|---|---|---|---|---|---|---|
| RTX 4090 desktop | cloth_shirt_4 |
4 transparent shirt cloth meshes, 50,944 cloth triangles | 10.961 | 11.058 | -0.9% | 0.856 | 3.063 | 64 |
| RTX 4090 desktop | shape_cloud_240 |
240 transparent primitive shapes plus opaque ground | 3.638 | 2.941 | +23.7% | 1.846 | 6.295 | 71 |
| RTX 4090 desktop | mesh_spheres_160 |
160 transparent generated sphere mesh instances | 1.846 | 1.357 | +36.0% | 0.994 | 4.386 | 66 |
| RTX 4090 desktop | bunny_meshes_24 |
24 transparent bunny.usd mesh instances, 292,800 mesh triangles | 2.002 | 1.484 | +34.9% | 0.360 | 3.112 | 93 |
| MacBook Pro M1 | cloth_shirt_4 |
4 transparent shirt cloth meshes, 50,944 cloth triangles | 13.541 | 16.152 | -16.2% | 0.839 | 3.049 | 66 |
| MacBook Pro M1 | shape_cloud_240 |
240 transparent primitive shapes plus opaque ground | 9.801 | 12.519 | -21.7% | 1.815 | 6.264 | 72 |
| MacBook Pro M1 | mesh_spheres_160 |
160 transparent generated sphere mesh instances | 8.889 | 12.313 | -27.8% | 0.971 | 4.367 | 66 |
| MacBook Pro M1 | bunny_meshes_24 |
24 transparent bunny.usd mesh instances, 292,800 mesh triangles | 8.190 | 11.386 | -28.1% | 0.356 | 3.101 | 98 |
| System | Scene | Description | Mean abs RGB | RMS RGB | Max abs RGB | Contact sheet |
|---|---|---|---|---|---|---|
| RTX 4090 desktop | franka_cloth |
Franka arm with transparent unisex_shirt.usd cloth |
0.183 | 1.190 | 46 | PNG |
| RTX 4090 desktop | h1_cloth |
H1 robot with transparent Style3D jacket cloth | 0.037 | 0.484 | 38 | PNG |
| RTX 4090 desktop | franka_robot |
Single transparent Franka arm | 0.042 | 0.570 | 69 | PNG |
| RTX 4090 desktop | franka_robot_grid |
3x3 grid of transparent Franka arms | 0.082 | 0.844 | 63 | PNG |
Data sources: RTX 4090
benchmark/results.json and
benchmark/summary.md; MacBook Pro M1
results.json,
summary.md, and
run-info.json; orbit
summary.md. The difference images multiply absolute
RGB differences to make subtle ordering/resolution differences visible.