Measuring Acceleration Structures
a year ago
- #performance
- #raytracing
- #GPU
- Hardware accelerated raytracing relies on acceleration structures (BVH/BLAS) for scene geometry, with varying memory layouts across vendors.
- Experimental setup uses Amazon Lumberyard Bistro scene, measuring BLAS sizes across different GPUs (AMD, NVIDIA, Intel).
- Results show significant variance in BLAS memory consumption, with NVIDIA leading efficiency (e.g., RTX 5070 at 18.8 bytes/triangle).
- AMD's RDNA4 introduces new BVH node formats (BVH8, primitive nodes) to reduce memory usage, achieving ~48 bytes/triangle.
- Driver optimizations (e.g., AMDVLK updates) have progressively improved BLAS memory efficiency over time.
- BVH structures balance memory and traversal performance, with trade-offs in node branching factors and triangle compression.
- Future improvements may further reduce BVH sizes, especially on newer GPU architectures like RDNA4 and NVIDIA's 5xxx series.