Research map / Point clouds / Point cloud classification
Flash3D: Super-scaling Point Transformers through Joint Hardware-Geometry Locality
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Flash3D is a point cloud transformer that uses Perfect Spatial Hashing to bucket points into GPU-tile-aligned memory blocks, enabling zero-overhead Swin-style shifting fused with FlashAttention-2, achieving 2.25x speedup and 2.4x memory savings over PTv3.
Problem
Point cloud transformer backbones like PTv3 rely on costly global serialization and shuffling of points that is misaligned with GPU memory tiling, limiting efficiency and scalability to larger attention scopes and models.
Approach
Use Perfect Spatial Hashing to scatter points into contiguous, GPU-tile-sized buckets so that geometric locality matches memory locality; subsequent attention rounds shift buckets logically (Bucket-and-Swin) instead of performing global reshuffles, fused with FlashAttention-2.
Key results
- On nuScenes at equal parameters (46.2M), Flash3D reaches 81.2 mIoU vs PTv3 80.4 with 2.25x faster latency (20ms vs 45ms on A100) and 2.4x less memory (0.5G vs 1.2G).
- With fixed memory quota, Flash3D scaled to 129.4M parameters reaches 81.5 mIoU, a 1.1-point gain over PTv3's 80.4.
- On ScanNet, Flash3D achieves 77.9 val / 78.6 test mIoU vs PTv3's 77.5 / 77.9, with A100 latency 24ms vs 61ms.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Flash3D instead of PTv3 at equal parameter count | point cloud semantic segmentation · nuScenes validation | mIoU: 80.4 -> 81.2 | Helped |
| Use Flash3D instead of PTv3 at equal parameter count | point cloud semantic segmentation · nuScenes on A100 | latency: 45ms -> 20ms | Helped |
| Use Flash3D instead of PTv3 at equal parameter count | point cloud semantic segmentation · nuScenes | memory: 1.2G -> 0.5G | Helped |
| Scale Flash3D to 129.4M parameters under a fixed memory quota | Flash3D · nuScenes validation | mIoU: 80.4 -> 81.5 | Helped |
| Use Flash3D instead of PTv3 at fixed memory quota | point cloud semantic segmentation · nuScenes on A100 | latency: 45ms -> 24ms | Helped |
| Use Flash3D instead of PTv3 | point cloud semantic segmentation · ScanNet Val | mIoU: 77.5 -> 77.9 | Helped |
| Use Flash3D instead of PTv3 | point cloud semantic segmentation · ScanNet Test | mIoU: 77.9 -> 78.6 | Helped |
| Add bucket rebalancing (ablation) | Flash3D with XD hash on nuScenes · nuScenes validation | mIoU: 78.6 -> 78.7 | Helped |
| Add bucket rebalancing (ablation) | Flash3D with XD+ZD hashes on nuScenes · nuScenes validation | mIoU: 79.1 -> 79.1 | No clear effect |
| Add bucket striding (ablation) | Flash3D with XD+ZD hashes on nuScenes · nuScenes validation | mIoU: 79.1 -> 79.8 | Helped |
| Add bucket shifting (Swin) (ablation) | Flash3D with XD+ZD hashes on nuScenes · nuScenes validation | mIoU: 79.1 -> 80.6 | Helped |
| Combine XOR-div with Zorder-div instead of using XOR-div alone (ablation) | Flash3D on nuScenes · nuScenes validation | mIoU: 78.6 -> 79.1 | Helped |
Surprising findings
- PSH latency is about two orders of magnitude lower than PTv3 serialization, and PTv3 uses under 5% TensorCore utilization in most scenarios.
- Bucket rebalancing has minimal effect on overall performance, and randomly perturbed bucket boundaries act as a regularizer.
Limitations
- Benchmarks limited to two datasets (ScanNet, nuScenes) in the main text; additional results deferred to appendix.
- Implementation depends on ThunderKittens and FlashAttention-2 CUDA kernels, with GPU-specific tuning.
- Results compare primarily against PTv3 with limited baselines.
Experimental setup
Datasets: ScanNet nuScenes
Benchmarks: ScanNet nuScenes
Metrics: mIoU latency memory parameters SM utilization TensorCore utilization DRAM bandwidth utilization
Baselines: PTv3 MinkUNet
Builds on: PTv3 FlashAttention-2 Swin Transformer Perfect Spatial Hashing
Closest papers by meaning
- Point Transformer V3: Simpler Faster Stronger (CVPR 2024 · similarity 0.88)
- FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision (NeurIPS 2024 · similarity 0.87)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.86)
- Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memory (NeurIPS 2025 · similarity 0.85)
- LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging (CVPR 2026 · similarity 0.85)
- FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.