Frame time is not steady in a static scene (rigged rebuilds, UI, input stalls, periodic work) #1625

Open
opened 2026-11-19 12:44:02 +00:00 by sky_line · 0 comments
Owner

Owner, 2026-11-19, RTX 5090: FPS should be rock steady in an unchanging scene, and it isn't.

  • 30-s FPS samples while standing still: 51, 57, 70, 71, 75, 86, 86, 84.
  • Small A/Bs (e.g. SSR's specialised shaders, 92 -> 89) can't be judged.

Hitch lines from the owner's log (a static scene, apart from settings toggles):

  • rw_buildRigged 25 ms: rigged geometry rebuilt with nothing changing.
  • vk_uiBuild / ui_draw 7-20 ms: UI build and draw, probably the Debug Settings floater, but UI cost should be flat.
  • df_gatherInput 40 ms: a main-thread stall.
  • vk_shadowRealloc 30 ms: an expected one-off on a settings change.

Goal: in a static scene, frame time is flat (low p99 jitter). Find every periodic or background contributor and remove it, or bound it, or spread it evenly:

  • CPU: rigged rebuilds, UI, input, texture / mesh streaming, eviction cycles, GC, logging and telemetry cadence, memory governor ticks;
  • GPU: adaptive passes toggling, history resets, async-queue contention, pacing.

Method:

  • A fixed-camera hold on the bench replay, on amdbox and garbage. Record per-frame CPU + GPU timers and the hitch attribution.
  • Report each contributor's period and cost. Include p50 / p95 / p99 and jitter before and after.

Part of the common-case optimisation work (RTX 3060-class, shadow maps; see agents/terra_telemetry_common_case_report.md).

Owner, 2026-11-19, RTX 5090: **FPS should be rock steady in an unchanging scene, and it isn't.** - 30-s FPS samples while standing still: 51, 57, 70, 71, 75, 86, 86, 84. - Small A/Bs (e.g. SSR's specialised shaders, 92 -> 89) can't be judged. **Hitch lines from the owner's log (a static scene, apart from settings toggles):** - `rw_buildRigged` 25 ms: rigged geometry rebuilt with nothing changing. - `vk_uiBuild` / `ui_draw` 7-20 ms: UI build and draw, probably the Debug Settings floater, but UI cost should be flat. - `df_gatherInput` 40 ms: a main-thread stall. - `vk_shadowRealloc` 30 ms: an expected one-off on a settings change. **Goal:** in a static scene, frame time is flat (low p99 jitter). Find every periodic or background contributor and remove it, or bound it, or spread it evenly: - CPU: rigged rebuilds, UI, input, texture / mesh streaming, eviction cycles, GC, logging and telemetry cadence, memory governor ticks; - GPU: adaptive passes toggling, history resets, async-queue contention, pacing. **Method:** - A fixed-camera hold on the bench replay, on amdbox and garbage. Record per-frame CPU + GPU timers and the hitch attribution. - Report each contributor's period and cost. Include p50 / p95 / p99 and jitter before and after. Part of the common-case optimisation work (RTX 3060-class, shadow maps; see agents/terra_telemetry_common_case_report.md).
sky_line added this to the Slipstream project 2026-11-19 12:44:02 +00:00
Sign in to join this conversation.
No milestone
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
sky_line/Slipstream#1625
No description provided.