Compressed BVH & Triangle Representations

Updated 30 June 2025

Compressed BVH and triangle representations are algorithms that use quantization (e.g., 10-bit encoding) and data packing to create compact spatial structures.
They accelerate spatial queries such as neighbor search, collision detection, and ray tracing by delivering 2–4× speedups and reducing redundant memory accesses.
These methods enable efficient storage by reducing triangle data by up to 16× and lowering overall bandwidth usage to as little as 18% of traditional implementations.

Compressed Bounding Volume Hierarchy (BVH) and Triangle Representations refer to families of algorithms, data structures, and encoding schemes that significantly reduce the memory footprint and improve traversal efficiency of BVHs and geometric primitives—chiefly triangles—for spatial queries in fields such as molecular simulation, collision detection, and ray tracing. Compression achieves substantial savings in bandwidth and storage while preserving correctness, and in some cases delivers additional acceleration in structural and algorithmic operations such as neighbor search, proximity queries, and intersection tests.

1. Quantized and Compressed BVH Structures

Early and foundational approaches, such as those developed for molecular simulation on GPUs, introduced quantized BVH nodes by discretizing space around the scene or simulation box into a regular grid, typically subdividing each axis into up to 1023 bins (for 10-bit quantization per axis) (1901.08088). Particle coordinates or bounding box bounds are encoded as 10-bit integers per axis, packed to minimize overhead. Axis-Aligned Bounding Boxes (AABBs) are then stored per node in this compact, quantized form. The construction follows steps:

Assign each particle or primitive to a spatial bin and compute a Morton code for spatial ordering.
Construct a linear BVH (LBVH) in parallel, with each node’s AABB bounds snapped to the quantization grid.
Store lower and upper bounds compactly; for three axes, the bounds are packed (fitting with pointers) into a 16-byte node.

Other methods employ reduced-precision representations, using half-precision floating-point (16 bits) for each BV descriptor rather than single-precision, and leverage SIMD instructions for on-the-fly decompression during traversal (2012.05348). More advanced tree layout strategies break the global BVH into "treelets"—small subtrees that fit within a CPU cache line—enabling cache-efficient traversal, lower memory traffic, and better parallelism.

2. Triangle and Primitive Compression

Compression of geometric primitives, especially triangles, is achieved by quantizing triangle vertex coordinates within the local coordinate system of a BVH node or meshlet. Algorithms quantize vertices to 8 or 10 bits per coordinate for each triangle relative to a grid defined by the BVH node, reducing triangle storage from 36 bytes (floats) to 9 bytes (quantized 8-bit integers) per triangle (2505.24653). In meshlet-based rendering, triangle indices are packed using optimal or near-optimal generalized triangle strips (GTSs), allowing 5–9 bits per triangle (after delta compression and bit-packing)—yielding index buffer compression up to 16:1 over the standard vertex pipeline (2404.06359).

For attribute data, crack-free vertex attribute quantization is enforced globally, so that shared vertices across meshlets map to identical quantized values and visual cracks are avoided (2404.06359). Meshlets store a local offset and scale for efficient decompression on GPU mesh shaders.

3. Traversal and Query Algorithms with Compression

Compressed BVH and triangle representations induce special considerations for traversal and spatial queries:

Neighbor and Proximity Search: In molecular simulation, stackless traversal visits nodes or particles whose quantized, conservatively expanded AABBs overlap a search volume (often a sphere) (1901.08088). All arithmetic is performed in single precision for speed, while quantization ensures no neighbors are missed, though some false positives are possible.
Collision Detection and Ray Tracing: Compressed BVHs support random access, simultaneous multi-ray traversal (enabling SIMD/AVX512 and ray stream tracing), and cache-aware layouts for throughput (2012.05348, 2505.24653). Traversal on fixed-point, quantized coordinates avoids floating-point rounding errors, enabling watertight intersection tests (no "holes") even with reduced precision.
Ray Stream Tracing: Multiple rays traversing the same BVH node are coalesced in a shared traversal stack, with each BVH node fetched only once for the group, minimizing redundant memory accesses and further reducing bandwidth (2505.24653).

4. Impacts on Memory, Performance, and Numerical Precision

Quantized and compressed data representations produce substantial reductions in both memory footprint and memory bandwidth consumption:

Node and Primitive Reduction: BVH nodes can be reduced to 16–96 bytes (depending on quantization and additional metadata), triangles to 5–9 bytes, and meshlet index buffers to less than 10% of previous representations (1901.08088, 2404.06359, 2505.24653).
Bandwidth Savings: Combined, these compressions lower memory traffic to as little as 18% of traditional implementations—crucial for bandwidth-limited environments such as mobile GPUs or hardware ray tracing units (2505.24653).
Traversal Speed: In molecular simulation, 2–4× speedup over state-of-the-art grid/cell-list neighbor searches is reported (1901.08088). In real-time rendering and large-scale scenes, compressed BVHs and stream tracing yield 5–10% improved traversal throughput solely from cache and data layout effects (2012.05348).
Numerical Robustness: Fixed-point arithmetic applied directly to quantized BVH and geometry ensures watertight intersection operations and eliminates floating-point rounding holes, provided all bounding/quantization steps are performed conservatively (2505.24653).

5. Algorithmic Enhancements: Heuristics and Predictive Coding

Compression is often improved through advanced coding schemes:

Surface Area Heuristic (SAH) and Distance-Aware BVHs: Incorporating cost models that weigh both surface area and spatial proximity of bounding volumes to rays leads to trees with tighter, task-specific node clustering. This reduces overlap, resulting in further reductions in invalid intersection tests and scene-specific acceleration—demonstrating performance improvements of 20–35% over traditional, area-only BVHs in channel modeling and ray tracing contexts (2208.10008).
Predictor-Corrector Encodings: Instead of encoding bounding boxes directly, the predictor-corrector scheme encodes each node as a delta from a predicted value (using its parent) (2012.05348). Deltas are typically small, enabling more aggressive quantization and storage as low as single bytes.
MILP-Based Strip Optimization: For meshlet compression, triangle orderings are optimized as generalized triangle strips using mixed integer linear programming (MILP), maximizing strip length and minimizing restart overhead (2404.06359).

6. Applications and Domain-Specific Adaptations

Compressed BVH and triangle representations are deployed in various high-performance and resource-limited contexts:

Molecular Simulation: For neighbor searches in molecular dynamics, quantized BVH outperforms grid (cell-list) approaches especially in low-density, high-inhomogeneity, or small search-cutoff scenarios (1901.08088).
Collision Detection and Robotics: Cache-sensitive, delta-compressed BVHs support interactive motion planning, real-time physics, and workspace monitoring (2012.05348).
Ray Tracing and Real-Time Rendering: Compressed and quantized BVHs, ray stream tracing, and meshlet compression are exploited in path tracing, mesh shading, and visualization of complex geometry on desktop and mobile devices (2404.06359, 2505.24653). Use cases extend to real-time graphics, virtual reality, and channel simulation for wireless networks (2208.10008).
Neural Hybrid Representations: Recent architectures integrate neural encodings within shallow BVHs, supporting neural ray queries with compression factors of 10–1000× over conventional geometry, all while integrating directly into standard rendering pipelines (2405.16237).

7. Practical Implementation Considerations and Limitations

Practical deployment of compressed BVH and triangle schemes entails several considerations:

Compression/Build Time: Adaptive or quantization-based builders may require longer to compute, but this cost is amortized over frequent query uses (1901.08088).
Memory–Precision Trade-offs: Increased compression may induce a small rise in false positives (e.g., up to ~5% for some simulations), coarser discretization, or increased per-node arithmetic. Conservative quantization strategies must be adopted to ensure correctness.
Random-Access and SIMD Suitability: Data layouts supporting treelet clustering and SIMD-accelerated decompress/traversal retain random access and branchless code paths (2012.05348).
Deployment: These techniques are especially suited to environments with constrained memory or limited bandwidth—such as embedded GPUs, mobile ray tracing, and custom hardware rasterizers (2505.24653). Opaque API and driver support are suggested as avenues to ease adoption.

Technique/Aspect	Compression Method	Key Benefit
BVH Node and AABB Representation	10-bit/8-bit quantization, half-float	2–4x node size reduction, bandwidth saving
Triangle/Primitive Representation	Local quantization, meshlet GTS	4–16x index and vertex buffer compression
Traversal/Query	Ray stream, cache-aware, SIMD	Up to 5–35% traversal speedup, 18% traffic
Watertightness/Precision	Fixed-point arithmetic	Geometric holes avoided, exact intersection
Heuristic Enhancement	SAH with spatial, predictor-corrector	Tighter bounds, further reduction in tests

Compressed BVH and triangle representations form a foundational component in memory- and bandwidth-efficient spatial query, rendering, simulation, and collision systems, enabling continued scalability and performance across diverse computational environments and application domains.