Benchmark Results: 3D Animation Tests
The 3D animation tests are currently centered on two products- Maxon’s Cinebench and LightWave 3D 9.6. This is partly due to availability and partly due to personal ability; we could create custom tests for LightWave and did so. We are also working on making versions of the tests for both Maya and 3ds Max.
The reported results are the average of three runs, and the tests are performed with Aero off, unless otherwise stated.
High Polygon Modeling Test
A very high polygon version of the Tom’s Hardware logo was created in LightWave, and this test duplicates one part of the process. A script is used to clone surface detailing across a part of the logo, and that script reports how long it took to run. This test isn't threaded, and the processor should enjoy the benefit of Turbo Boost while it is running.
Note that with Hyper-Threading enabled, the results get worse, not better. This is just the first time we will see this, as the pattern is repeated later on.
High Polygon Rendering Test
This test consists of the Tom’s Hardware logo in its 1.68 million polygon-glory, with four example frames rendered from a 600 frame animation. The polygon count was deliberately targeted, as it is close to the number of polygons in a ‘hero model’ level of detail for television animation. As an example, the Battlestar Pegasus from the recent Battlestar Galactica TV series was about the same polygon count, but with vastly more texture maps. The rendering itself makes use of anti-aliasing (nine passes), photorealistic motion blur, ray traced shadows, and global illumination. Once again, these render settings are considered the norm in TV visual effects. The frames are rendered at 1920x1080.
Unlike the modeling test, the render test is very well threaded. On the average, a 10%+ performance gain is realized while rendering with multithreading on. The different frames of the sequence represent different problems for the renderer, and frame 500 shows the largest render time because it has the maximum number of polygons visible at the largest size on screen at the same time. In frame 600, meanwhile, the camera is at its greatest distance and stationary, so the frame differences for motion blur calculation become relatively small.
OpenGL Preview Test
Here we're measuring the time required to generate an OpenGL preview of the entire 600-frame animation for real-time playback at the animation’s actual frame rate of 24 FPS.
Note that this machine can’t push our model around in real-time. The animation was done with a low-resolution proxy object, and that was replaced with the high polygon object for the final lighting and rendering. This test is performed at both 1680x1050 and 1920x1080.
Once again, a process that is not well-threaded actually runs faster with Hyper-Threading turned off. Even the best time (33.32 seconds at 1680x1050) is 8.32 seconds longer than real-time. That should give you an idea of how high the system load is with this scene.