Skip to content

Acceleration checkpoints

Updated 7 November 2026. The physical MBX performance observations below retain their original dates and workloads.

Component Recorded result Work remaining
RGB console Twelve exact RGB565 swatches; normal boot verified on 16 October. Continued console regression coverage.
C cell compositor Loaded and used by an early N810 application; retained buffers and changed-span output measured. Whole-desktop responsiveness across workloads.
ARM VFP Physical numerical tests and later Mesa native floating-point execution. Cost reduction across the complete frame pipeline.
C55x DSP Assembler, encoding checks, and numerical models. Physical loading, execution, and output verification.
MBX rasterization Standard 400 by 400 glxgears rendered on the Nokia. Broader fixed-function state support and production integration.
VGP geometry Diagnostic transforms and lighting compared with Mesa references; live GPU geometry in the 4 November gears runs. Broader shader and geometry coverage.
Display transfer Approximately 16.66 DISPC interrupts per second in the earlier comparison and 23.4 in the later GPU-geometry trials. RFBI/DISPC crop and transfer investigation.
Native graphical desktop Shared desktop model, direct Room pixels, and actual OS resize/input checks on Apple M1 Pro; source builds and explicit software-GL checks in Linux containers. Physical Nokia frontend and mixed-state rendering acceptance.
Mac GPU effects and maps Exact-output comparisons for Flux, Flurry, windows, DEM shading, and terrain; 738-frame combined experimental run. Matched timing comparisons, remaining CPU preparation/repairs, and session-save notification follow-up.
Symbian GPU drawing Physical checks of selected clears, retained surfaces, scissor rectangles, and large image sampling. Full GPU-backed UI and interaction for the test applications.

The native frontend guide describes the included source and build commands. The Mac GPU journal lists the tested stages and CPU work retained. The Symbian surface journal records experimental driver results and failed attempts.

Selected performance observations

Workload Observation Measurement scope
Water scene, 20 October 7.141 s geometry; 65.218 s to coherent app output. One attended N810 launch.
Saved whole gears frame, 3 November 0.118132 s for one GPU job and 390 primitives. Offscreen saved-frame execution.
Native VFP gears checkpoint, 3 November 29.2825 FPS median after two startup intervals. Live application intervals; a slow interval reached 12.373 FPS.
Quiet control, 4 November 30.7405 FPS median; 28.3327 completed GPU FPS across the measured window. Pocket UI temporarily paused during the bounded run.
Quiet optimized, 4 November 33.596 FPS median; 30.9953 completed GPU FPS across the measured window. Same application, dimensions, and quality; Pocket UI resumed afterward.

The later GPU-geometry checkpoint completed 1,864 frames in 60.711 seconds, or 30.70 FPS across two repeat intervals. Application CPU cost was 17.65 ms per frame. The same-module polling-disabled control measured 28.77 FPS and 18.24 ms per frame. The journal's repeated comparison includes the physical framebuffer, individual runs, failed alternatives, and downloadable measurements.

GPU completion, application-reported FPS, and physical display transfers measure different stages. The performance entry explains the background-load and profiling effects.