poulscube turns the project’s cube logo into a small OpenGL workload. Its few draws exercise colored geometry, perspective, depth and the path from Mesa to an X11 window. In the capture build, vertex shading and rasterization run on SGX535, and X11 uses the 2D core to copy completed windowed frames.
The image above is a frame read from the GPU’s back buffer during the capture step. The capture clock is fixed to make image comparisons repeatable.
Measured on the Dell
| Window | FPS | Date | Measurement |
|---|---|---|---|
| 300 × 300 | 399.4 | 2026-11-03 | Median of three repeats |
| 640 × 640 | 117.5 | 2026-11-03 | Median of three repeats |
- Platform
- GMA500 · Debian 12 i686
- Governor / swap interval
- ondemand / 0
- Measured build
5c985afe34d2
The measured run completed without GPU faults. Work on CPU overhead continues.
Median of post-startup five-second FPS samples in each 21-second run, then the median of three runs per size. Synchronization is disabled.
Rendering status
Selected GMA500 tests pass for small framebuffer objects, changing frame layouts and switching between differently sized targets. Full API and FBO coverage remain in progress.
What changed
These results follow GPU vertex shading, overlap between X11’s 2D copies and 3D rasterization, and later overlap between one scene’s geometry work and another scene’s raster work. Giving the GPU room to work explains the earlier 2D-copy/raster overlap experiment and the memory contention it revealed.