Rendering and loading performance

Version 0.1.0 coalesces model-loading visibility, clipping, selection and diagnostic refreshes into the existing draw scheduler. Empty selections skip pickable-object matching and bounding-box traversal. Close-zoom labels reuse transformed pad positions and a projection vector. No additional animation loop or polling timer is introduced by this optimization.

Browser sample

The credited OLIMEX ESP32-P4-DevKit Rev D1 demo loaded 141 model components from local files, in the same desktop Chromium browser and viewport. The instrumented previous renderer (source revision cf3b59f) performed 141 model refreshes and spent approximately 342 ms in visibility, clipping, selection and diagnostic work. The candidate performed 9 batched refreshes and spent approximately 10 ms in that work. A later candidate remount performed 7 refreshes in approximately 12 ms; scheduling and cache state affect these numbers.

These are bounded loading-work samples, not an FPS, network, total-load-time or cross-device benchmark. The instrumentation excludes model fetch and decode time. The baseline bundle had observation counters added; the candidate exposes counters through getRenderStats().

Translucent inspection cost

Internal copper geometry is created lazily on the first translucent view and reused across appearance changes. Materials preserve original assets and are restored on return to an opaque appearance. Extra visible internal copper adds geometry and draw calls; translucency is an inspection feature, not a promise of a faster render mode. Screen-space ambient occlusion and object shadows are disabled while translucent. Use Low or Balanced quality for smaller devices.

The source XY geometry and layer identity are retained. Internal elevations use a complete consistent saved stackup where available; otherwise the viewer reports illustrative layer spacing. Blind/buried via geometry and unsupported constructs remain covered by rendering diagnostics. Physical mobile GPU performance has not been measured.

0.2.1 mobile embed sample

A 340-pixel-wide embed of the credited OLIMEX ESP32-P4-DevKit Rev.D1 board was checked in the same desktop browser, using Balanced rendering and proxy component models. Three samples gave median source-to-PCB load times of 507.7 ms before and 296.5 ms after, and median 3D mount times of 1263.3 ms before and 736.0 ms after (about 42% lower in this sample). Source download was excluded; this does not measure physical-phone FPS, total network loading, battery use or peak memory.

Twenty separated display updates while the embed was off-screen produced 20 draws before and zero after, followed by one coalesced refresh on return. Selecting the sampled signal net avoided 160 bounds evaluations with zero bounding-box or outline-buffer work; selection identity remained intact. Unchanged file menus and theme icons retained their nodes across 100 control updates. See mobile website integration for loading and lifecycle guidance.

1.0.0: overlapping copper stability

Release 1.0.0 disables hardware multisample depth for the logarithmic-depth board renderer. Balanced, Photoreal Fast, High and Presentation smooth resolved color with SMAA; Low renders directly. Copper cap precedence is applied in the logarithmic depth domain and survives selection/translucency material clones. Source coordinates and geometry remain unchanged. Balanced uses at most two million pixels (one million on constrained devices); Photoreal Fast uses at most one million. The conservative allocation estimate includes the added color targets and retains the fast preset’s existing 60 MiB framebuffer ceiling. These are allocation bounds, not physical-phone timing results.

Next release: repeatable stability measurements

The unreleased stability demo reports viewer-ready timing, 20 synchronous linked-net select/clear samples and renderer-owned resource estimates for KERO/IRCA. It also asserts that net selection leaves the 3D camera and expansion unchanged. See the validation record for samples and limits; these are not physical-phone FPS, peak memory or before/after speedup claims.

Next release: source and repeated-assembly retention

Compact prepared source, shared dim materials and cross-layer pad buffers reduce avoidable retained allocations. See memory optimization for the 100,000-object synthetic measurements, repeated-assembly tests and scope. The prepared-source stage saved 29.6% retained heap in the three-run sample; this is not total browser/GPU RAM or phone performance.