Optimizations of voxel renderers are documented mostly as engineering advice
without measurements. In our survey of the field, we found no comparison
that measures several techniques on the same data and in the same system.
In this thesis we measure several such techniques on a demanding, real-world
workload. We display the entire Slovenian national LiDAR survey published
as of December 2025, almost three quarters of the country, in a web browser
as voxel terrain using the WebGPU API, with no render-distance limit. We
condense the point cloud into heightmaps of five bytes per cell, build a levelof-detail pyramid of delta chunks from them (74.8 GB, only 1.6% more than
the base data), and stream it over a WebSocket connection. Into the same
system we integrate six rendering approaches (an explicit mesh, instanced
cubes, instanced faces, greedy-meshed faces, heightmap ray casting, and a
hybrid approach) and compare them in terms of speed, memory use, and
data transfer with an automated benchmarking harness, and in terms of
implementation difficulty by lines of code together with a qualitative rating
of how hard each approach was to get right. Four of the approaches reach
the goal of 60 frames per second when viewing all of the data at once. Ray
casting and the hybrid approach are the fastest by median at roughly 390
frames per second, and for such a view we download only 45 MB of data. We
distil the results into recommendations for voxel-engine developers.
|
