RT VRAM: hardware compressed / cluster acceleration structures (NVIDIA) #581

New issue

Open

opened 2026-10-09 01:49:13 +00:00 by sky_line · 0 comments

sky_line commented

2026-10-09 01:49:13 +00:00

Owner

Measured in a heavy region (2026-10-09): RT acceleration structures (rt-as) are about 1.7 GB, as much as all the geometry, and the opacity micromaps (rt-mask) are 354 MB. Follow-up to #573.

Idea

Use NVIDIA cluster acceleration structures (RTX Mega Geometry, VK_NV_cluster_acceleration_structure) and/or other compressed-BLAS extensions where available.

Trade-offs

Large memory savings with no visual loss.
NVIDIA-only: it needs a separate code path, feature detection, and the current BLAS path kept as the fallback for AMD, Intel and older NVIDIA. That's a maintenance burden.

Done when

On supported cards, rt-as drops by a large factor with no change in output, and other cards are unaffected.

Measured in a heavy region (2026-10-09): RT acceleration structures (`rt-as`) are about 1.7 GB, as much as all the geometry, and the opacity micromaps (`rt-mask`) are 354 MB. Follow-up to #573. ## Idea Use NVIDIA cluster acceleration structures (RTX Mega Geometry, VK_NV_cluster_acceleration_structure) and/or other compressed-BLAS extensions where available. ## Trade-offs - Large memory savings with no visual loss. - NVIDIA-only: it needs a separate code path, feature detection, and the current BLAS path kept as the fallback for AMD, Intel and older NVIDIA. That's a maintenance burden. ## Done when On supported cards, `rt-as` drops by a large factor with no change in output, and other cards are unaffected.

sky_line added the

rendering

perf

triage

labels

2026-10-09 01:49:13 +00:00

sky_line added this to the Slipstream project

2026-10-09 01:49:14 +00:00

sky_line referenced this issue

2026-10-09 04:40:41 +00:00

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

sky_line added