diff --git a/implementations/Crafter.Graphics-RenderingElement3D-WebGPU.cpp b/implementations/Crafter.Graphics-RenderingElement3D-WebGPU.cpp
index 326f049..2cafd67 100644
--- a/implementations/Crafter.Graphics-RenderingElement3D-WebGPU.cpp
+++ b/implementations/Crafter.Graphics-RenderingElement3D-WebGPU.cpp
@@ -159,7 +159,11 @@ void RenderingElement3D::BuildTLASUpload(WebGPUCommandEncoderRef /*cmd*/, std::u
         runOwned = currOwned;
     }
 
-    tlas.metadataBuffer.FlushDevice();
+    // Only primitiveCount slots are live; the buffer is padded to kNPadded
+    // (65536). Flushing the whole thing wgpuWriteBuffers all 256 KB through
+    // the WASM->JS staging path every frame — ~100-250x waste for a
+    // few-hundred-instance scene. Range-flush just the live entries.
+    tlas.metadataBuffer.FlushDeviceRange(0, 0, primitiveCount * sizeof(std::uint32_t));
 }
 
 void RenderingElement3D::BuildTLASBuild(WebGPUCommandEncoderRef /*cmd*/, std::uint32_t index) {