FlashFPS: Efficient Farthest Point Sampling for Large-Scale Point Clouds via Pruning and Caching

2026-04-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

FlashFPS is a novel, hardware-agnostic framework designed to accelerate Farthest Point Sampling (FPS) in Point-based Neural Networks (PNNs), a critical operation that often causes significant inference latency, especially for large-scale point clouds. The framework, released on April 20, 2026, addresses three key redundancies in FPS: full-cloud computations, late-stage iterations, and predictable inter-layer outputs. FlashFPS integrates two main components: FPS-Prune, which uses candidate and iteration pruning to reduce redundant computations while maintaining sampling quality, and FPS-Cache, which eliminates layer-wise redundancy through caching and reuse. When integrated into existing CUDA libraries and PNN accelerators, FlashFPS achieves a 5.16x speedup over the standard CUDA baseline on GPU and a 2.69x speedup on PNN accelerators, with minimal accuracy loss, thereby enabling more efficient and scalable PNN inference.

Key takeaway

For AI Engineers optimizing point cloud processing, FlashFPS offers a significant performance boost for Farthest Point Sampling. You should consider integrating this plug-and-play framework into your PNN pipelines to achieve substantial speedups, up to 5.16x on GPUs, without compromising accuracy, thereby enhancing the scalability of your models.

Key insights

FlashFPS accelerates point cloud processing by pruning and caching Farthest Point Sampling computations.

Principles

• Identify and eliminate computational redundancies.
• Preserve quality while optimizing for speed.
• Leverage inter-layer predictability for efficiency.

Method

FlashFPS employs FPS-Prune for candidate and iteration pruning, and FPS-Cache for layer-wise cache-and-reuse, reducing redundant computations in Farthest Point Sampling.

In practice

• Integrate FlashFPS into CUDA libraries.
• Apply FlashFPS to PNN accelerators.
• Utilize pruning for large-scale point clouds.

Topics

• Farthest Point Sampling
• Point Cloud Processing
• Point-based Neural Networks
• FlashFPS Framework
• Computational Efficiency

Code references

• Yuzhe-Fu/FlashFPS
• auniquesun/Point-Cache
• KonradSzafer/Pointy

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Learned Radius Estimation for UDF-Based Point Cloud Reconstruction — Learning to adapt the support radius dynamically improves UDF-based point cloud reconstruction accuracy, especially for complex geometries.
• HASTE: Hardware-Aware Dynamic Sparse Training for Large Output Spaces — HASTE optimizes XMC output layers via group-shared sparsity and a dense/sparse decomposition for hardware-aware speedups.
• PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain — Improved data quality and multi-token latent spaces enable stable, high-performance, low-latency 3D scene completion.
• Synthetic Benchmarks Overstate Forward-Forward Scaling: Real-Data Limits of Layer-Local Training — Layer-local Forward-Forward training, even at SOTA, struggles to scale to real-world data and offers no practical memory advantage over optimized backpropagation.
• State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives — Wavelet-based DPPs and a novel discretization pipeline significantly enhance minibatch and coreset sampling efficiency for diverse ML tasks.
• Federated Gaussian Process Learning via Pseudo-Representations for Large-Scale Multi-Robot Systems — pxpGP enables scalable, privacy-preserving Gaussian Process learning for large multi-robot systems via pseudo-representations and optimized ADMM.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.