perf(renderer): investigate ~2.3 ms GPU volumetric cost on Windows laptop #1080

Closed
opened 2026-11-07 03:28:22 +00:00 by SakulFlee · 0 comments
Owner

Summary

The Windows laptop capture is GPU-bound: CPU is around 5% and GPU is above 90%. Taking the GPU snapshot at face value:

  • gpu_volumetrics: ~2.27 ms mean (2.06-2.41)
  • gpu_models: ~1.80 ms
  • gpu_shadow: ~1.28 ms
  • gpu_tonemap: ~0.87 ms
  • gpu_aa: ~0.21 ms
  • gpu_scene_total: ~3.16 ms
  • Aggregate: ~6.51 ms, versus a ~6.46 ms frame

Volumetrics is the largest single GPU stage and a plausible target for the RTX 3050 laptop.

Blocker

The GPU timestamp readback currently goes stale (#1076), so these numbers must be confirmed after fixing that. Resolution, refresh, and volumetric quality are also missing from the capture (#1078).

Investigation

After timestamp readback is trustworthy:

  • A/B volumetrics on/off at the same camera and resolution.
  • Measure froxel grid resolution, composite resolution, quality preset, and render scale.
  • Check whether the cost scales with resolution or with light/froxel count.
  • Compare Windows and Linux at matched resolution and quality.
  • Decide whether Windows needs adaptive volumetric quality or whether ~2.3 ms is expected for this GPU at the capture resolution.

Acceptance criteria

  • A trustworthy per-pass GPU breakdown is attached.
  • The dominant volumetric cost is identified.
  • Either an optimization issue is created with measured targets, or the cost is documented as expected.
## Summary The Windows laptop capture is GPU-bound: CPU is around 5% and GPU is above 90%. Taking the GPU snapshot at face value: - `gpu_volumetrics`: ~2.27 ms mean (2.06-2.41) - `gpu_models`: ~1.80 ms - `gpu_shadow`: ~1.28 ms - `gpu_tonemap`: ~0.87 ms - `gpu_aa`: ~0.21 ms - `gpu_scene_total`: ~3.16 ms - Aggregate: ~6.51 ms, versus a ~6.46 ms frame Volumetrics is the largest single GPU stage and a plausible target for the RTX 3050 laptop. ## Blocker The GPU timestamp readback currently goes stale (#1076), so these numbers must be confirmed after fixing that. Resolution, refresh, and volumetric quality are also missing from the capture (#1078). ## Investigation After timestamp readback is trustworthy: - A/B volumetrics on/off at the same camera and resolution. - Measure froxel grid resolution, composite resolution, quality preset, and render scale. - Check whether the cost scales with resolution or with light/froxel count. - Compare Windows and Linux at matched resolution and quality. - Decide whether Windows needs adaptive volumetric quality or whether ~2.3 ms is expected for this GPU at the capture resolution. ## Acceptance criteria - A trustworthy per-pass GPU breakdown is attached. - The dominant volumetric cost is identified. - Either an optimization issue is created with measured targets, or the cost is documented as expected.
Sign in to join this conversation.
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
SakulFlee/Orbital#1080
No description provided.