bevyengine · alice-i-cecile · Aug 14, 2025 · Aug 11, 2025 · Aug 12, 2025 · Aug 12, 2025
diff --git a/release-content/release-notes/meshlet-bvh-culling.md b/release-content/release-notes/meshlet-bvh-culling.md
@@ -0,0 +1,19 @@
+---
+title: Meshlet BVH culling
+authors: ["@SparkyPotato", "@atlv24"]
+pull_requests: [19318]
+---
+
+(TODO: Embed example screenshot here)
+
+Bevy's virtual geometry has been greatly optimized with BVH-based culling, leading to almost true scene-complexity invariance on the GPU.
-Bevy's virtual geometry has been greatly optimized with BVH-based culling, leading to almost true scene-complexity invariance on the GPU.
+Bevy's virtual geometry has been greatly optimized with BVH-based culling, making the cost of rendering nearly independent of scene geometry. (Then write something here about 120k vs 1 million instances).
-Bevy's virtual geometry has been greatly optimized with BVH-based culling, leading to almost true scene-complexity invariance on the GPU.
+Bevy's virtual geometry has been greatly optimized with BVH-based culling, making the cost of rendering nearly independent of scene geometry. (Then write something here about 120k vs 1 million instances).
+
+This gets rid of the previous cluster limit that limited the world to 2^24 clusters (about 4 billion triangles).
-This gets rid of the previous cluster limit that limited the world to 2^24 clusters (about 4 billion triangles).
+This also gets rid of the previous cluster limit that limited the world to 2^24 clusters (about 4 billion triangles).
-This gets rid of the previous cluster limit that limited the world to 2^24 clusters (about 4 billion triangles).
+This also gets rid of the previous cluster limit that limited the world to 2^24 clusters (about 4 billion triangles).
+There are now *no* hardcoded limits to scene size, only unique instance limits due to VRAM usage (since streaming is not yet implemented),
+and total instance limits due the current architecture requiring all instances to be uploaded to the GPU every frame.
-There are now *no* hardcoded limits to scene size, only unique instance limits due to VRAM usage (since streaming is not yet implemented),
-and total instance limits due the current architecture requiring all instances to be uploaded to the GPU every frame.
+There are now *no* hardcoded limits to scene size. In practice you will only be limited by asset VRAM usage (since streaming is not yet implemented),
+and total instance count due the current code requiring all instances to be re-uploaded to the GPU every frame.
-There are now *no* hardcoded limits to scene size, only unique instance limits due to VRAM usage (since streaming is not yet implemented),
-and total instance limits due the current architecture requiring all instances to be uploaded to the GPU every frame.
+There are now *no* hardcoded limits to scene size. In practice you will only be limited by asset VRAM usage (since streaming is not yet implemented),
+and total instance count due the current code requiring all instances to be re-uploaded to the GPU every frame.
+
+The screenshot above has 130,000 dragons in the scene, each with about 870,000 triangles, leading to over *115 billion* total triangles in the scene.
+However, this still runs at 60 fps on an RTX 4070 at 1440p, with most of the time being due to the instance upload CPU bottleneck mentioned above (taking 14 ms of CPU time).
+
+Speaking of GPU cost, the scene above renders in about 3.5 ms on the 4070, with ~3.1 ms being spent on the geometry render and ~0.4 ms on the material evaluation.
-Speaking of GPU cost, the scene above renders in about 3.5 ms on the 4070, with ~3.1 ms being spent on the geometry render and ~0.4 ms on the material evaluation.
+Speaking of GPU cost, the scene above renders in about 3.5 ms on a 4070, with ~3.1 ms being spent on the geometry render and ~0.4 ms on the material evaluation.
-Speaking of GPU cost, the scene above renders in about 3.5 ms on the 4070, with ~3.1 ms being spent on the geometry render and ~0.4 ms on the material evaluation.
+Speaking of GPU cost, the scene above renders in about 3.5 ms on a 4070, with ~3.1 ms being spent on the geometry render and ~0.4 ms on the material evaluation.
+Increasing the instance count to over 1 million (almost *900 billion triangles*!), the total increase to about 4.5 ms, with ~4.1 ms on geometry render and material evaluation remaining constant at ~0.4 ms.