Those working with point cloud processing are likely familiar with PDAL. It is open source and easy to use, but when dealing with large-scale point cloud data, the original PDAL relies purely on the CPU and takes a long time. Recently, an open-source project called GPUPDAL appeared on GitHub, using GPUs for compute-intensive tasks. In actual tests, it is 2 to 14 times faster than the original. Today I will walk you through it.

Project Overview

Many people refer to PDAL as "the GDAL of point clouds", and the two are indeed closely related. GDAL unified the abstract reading and writing of raster and vector data. Later, Howard Butler, one of GDAL's core contributors, created another project following this architectural approach: PDAL (Point Data Abstraction Library), specifically for processing 3D point clouds.

GPUPDAL (GPU Point Data Abstraction Library) is a high-performance branch built on PDAL by developer Zy Mazza. Its core positioning is a seamless drop-in replacement for the pdal command line: the original pipeline command and JSON configuration do not need to be changed; simply switch to gpupdal and it runs directly. It also provides compatibility: when encountering unsupported operators or small-data scenarios, it automatically falls back to the original CPU implementation.

PDAL open-source repository: https://github.com/PDAL/PDAL

GPUPDAL open-source repository: https://github.com/zymazza/GPUPDAL

Performance

GPUPDAL's speedup comes from two aspects. First, CUDA kernels take over compute-intensive stages such as ground point classification and normal vector computation. Second, memory mapping and multithreaded I/O have been rewritten. Even without GPU enabled, a single test converting a 22-million-point LAS 1.4 file was more than ten times faster.

The official time comparison for 18 common tasks is as follows:

On a workstation, GPUPDAL achieved an overall speedup of about 2x to 14x across this task set. The following image shows the DTM hillshade output after GPUPDAL processed a 47.5-million-point AHN4 tile in the Netherlands:

Installation and Quick Start

The project can be installed globally on Linux and Windows (it can be installed even without an NVIDIA GPU, and will automatically fall back to CPU mode):

npm install --global gpupdal
gpupdal --version

Daily usage is the same as the original pdal:

gpupdal pipeline pipeline.json
gpupdal translate input.laz output.laz
gpupdal doctor

At present, the project has several installation prerequisites:

  • GPU ecosystem: currently only NVIDIA CUDA is supported, requiring driver 580 or later (source compilation requires CUDA 12.4+).
  • Support scope: at this stage, full automatic acceleration validation mainly targets RTX 4090 (SM 8.9 architecture) on Linux; on Windows, exact execution has only been validated on NVIDIA L4.
  • Streamlined plugins: the first binary release disables some optional external plugins by default, retaining only core driver directories.

After installing on a new machine, it is recommended to run gpupdal calibrate first. The tool measures the actual CPU and GPU execution times on the local machine and generates a local scheduling profile at ~/.config/gpupdal/placement-profile.json, so that small files do not become slower due to GPU transfer overhead. Before replacing production scripts, you can also run gpupdal verify to compare against the original pdal on the same machine and directly generate JSON and HTML verification reports.

Final Thoughts

Point cloud datasets are getting larger. In the past, the time cost of relying solely on CPU for batch processing was indeed very high. GPUPDAL's approach of offloading high-load operators to the GPU while retaining the original exact fallback mechanism is a valuable reference for LiDAR preprocessing workflows. However, the project is still in an early iteration stage, is tightly tied to NVIDIA hardware, and has some plugins disabled by default. Before replacing it in production, it is recommended to test it with real data.

Has anyone used GPUPDAL? How did it perform? Feel free to share your thoughts in the comments.

References

  1. GPUPDAL GitHub repository: https://github.com/zymazza/GPUPDAL
  2. Hacker News discussion (Show HN: GPU Accelerated PDAL): https://news.ycombinator.com/item?id=49435679