Epic: VRAM pressure redesign -- texture pool absorbs pressure first, RT fits 8 GB cards, no alarming notices

sky_line commented

2026-10-04 05:59:13 +00:00

Owner

Users are confused by "Graphics card memory is nearly full" notices, and RT shadows drop out on 8 GB cards. Goal: VRAM pressure is absorbed silently, mostly by the texture pool, and RT fits 8 GB cards.

Done (shipped via the #405 budget manager, `llvrambudgetvk.cpp`)

Silent soft ladder: over-demand textures -> off-screen textures -> far textures lose GPU mips first, driven by per-texture renderer demand; global mip bias is a late rung.
Ladder order (owner, 2026-10-05) is the spec: probes, mesh LOD, far shadows, far draw distance, RT budget (geometry budget -> avatars out of RT -> RT shadow resolution 3/4, 1/2 -> RT range 64 m) and mirrors come before near textures.
Hard step (last resort only): settings lowered + latched, one plain notification each; soft rungs only log.
BLAS compaction; recovery hysteresis on the way back up; #164 (RT inside a VRAM-scaled budget) closed.

Remaining: 8 GB fit

First-launch defaults by VRAM tier (<= 8 GB: RT shadow mask Half, RT radius 128 m, probes 128 px x 64, BC compression on). No tier code exists today; everyone gets the same defaults.
Smaller RT shadow mask: ~10 images incl. RGBA32F depth and several RGBA16F (acc/meta/history x2); move to smaller formats / fewer images. Today only the mask scale drops under pressure.
Trim owned pools under pressure: texture staging ring and idle CPU texture copies (only encoder pools are released today).
Other apps' VRAM only slowly: osSmoothed drops at once when another app allocates; make the downward reaction gradual too (still fast enough to avoid device-lost).
Confirm with an 8 GB user log (VKMem lines) once the above land.

RT geometry memory is tracked separately: #578 (lower-LOD BLAS for distant objects), #579 (drop tiny distant objects from the TLAS), #580 (lower opacity-micromap resolution far away), #581 (compressed / cluster acceleration structures).

Part of #405.

Users are confused by "Graphics card memory is nearly full" notices, and RT shadows drop out on 8 GB cards. Goal: VRAM pressure is absorbed silently, mostly by the texture pool, and RT fits 8 GB cards. ### Done (shipped via the #405 budget manager, `llvrambudgetvk.cpp`) - Silent soft ladder: over-demand textures -> off-screen textures -> far textures lose GPU mips first, driven by per-texture renderer demand; global mip bias is a late rung. - Ladder order (owner, 2026-10-05) is the spec: probes, mesh LOD, far shadows, far draw distance, RT budget (geometry budget -> avatars out of RT -> RT shadow resolution 3/4, 1/2 -> RT range 64 m) and mirrors come **before** near textures. - Hard step (last resort only): settings lowered + latched, one plain notification each; soft rungs only log. - BLAS compaction; recovery hysteresis on the way back up; #164 (RT inside a VRAM-scaled budget) closed. ### Remaining: 8 GB fit - [ ] **First-launch defaults by VRAM tier** (<= 8 GB: RT shadow mask Half, RT radius 128 m, probes 128 px x 64, BC compression on). No tier code exists today; everyone gets the same defaults. - [ ] **Smaller RT shadow mask**: ~10 images incl. RGBA32F depth and several RGBA16F (acc/meta/history x2); move to smaller formats / fewer images. Today only the mask scale drops under pressure. - [ ] **Trim owned pools under pressure**: texture staging ring and idle CPU texture copies (only encoder pools are released today). - [ ] **Other apps' VRAM only slowly**: `osSmoothed` drops at once when another app allocates; make the downward reaction gradual too (still fast enough to avoid device-lost). - [ ] Confirm with an 8 GB user log (VKMem lines) once the above land. RT geometry memory is tracked separately: #578 (lower-LOD BLAS for distant objects), #579 (drop tiny distant objects from the TLAS), #580 (lower opacity-micromap resolution far away), #581 (compressed / cluster acceleration structures). Part of #405.

sky_line added the

epic

perf

rendering

labels

2026-10-04 05:59:13 +00:00

sky_line referenced this issue from a commit

2026-10-04 06:45:16 +00:00

#282 phase 1: VRAM pressure ladder -- texture pool absorbs pressure first, silently

sky_line referenced this issue from a commit

2026-10-04 06:45:16 +00:00

#282: one texture-step owner -- budget fit is demand-sorted too, global bias last

sky_line referenced this issue

2026-10-04 12:58:51 +00:00

VRAM pressure: reflection probes as a ladder step (resolution, then count), probe budget, no hard OOM aborts (#294) #295

sky_line referenced this issue

2026-10-04 13:06:27 +00:00

Explore RTX Mega Geometry: partitioned TLAS (PTLAS) and cluster BLAS (CLAS), measure first #296

sky_line referenced this issue

2026-10-05 03:08:10 +00:00

Epic: RAM/VRAM management -- one budget, tiered texture storage, run on a stated minimum #405

sky_line commented

2026-10-05 03:08:10 +00:00

Author

Owner

Part of #405 (RAM/VRAM management epic).

sky_line added a new dependency

2026-10-05 03:08:10 +00:00

#405 Epic: RAM/VRAM management -- one budget, tiered texture storage, run on a stated minimum

sky_line referenced this issue from a commit

2026-10-05 08:37:46 +00:00

VRAM ladder wired: effective caps, no setting writes under ordinary pressure; hard step last (#405, #227, #226)

sky_line added this to the Slipstream project

2026-10-06 15:04:14 +00:00

sky_line commented

2026-10-09 08:53:46 +00:00

Author

Owner

Rewritten to match current state: Phase 1 shipped via the #405 budget manager (with the owner's 2026-10-05 ladder order as the spec); what remains is the 8 GB-fit list.

sky_line referenced this issue

2026-10-09 09:20:39 +00:00

VRAM: allocate the projector shadow atlas (and sun cascades) only when shadow maps are actually used -- ~320 MB back for ray-traced shadows #600

sky_line referenced this issue

2026-10-10 04:53:22 +00:00

Epic: RAM/VRAM management -- one budget, tiered texture storage, run on a stated minimum #405

sky_line referenced this issue

2026-10-19 16:32:07 +00:00

Snapshots on the main render path: grab the live frame (downscale) or jittered accumulation (upscale) instead of an offscreen re-render #873