Unexpectedly slower performance on RTX 4060 vs README's GTX 1050 Ti numbers

[Zyl225]

Hi, thanks for open-sourcing this great CUDA voxelizer.

I tried to reproduce the performance numbers from the README and noticed that my results on a newer GPU are significantly slower than what is reported for a GTX 1050 Ti.

Environment

• GPU: NVIDIA GeForce RTX 4060
• CUDA: 12.1
• OS: Windows 10
• Driver version: [e.g. 555.xx]
• Build: [prebuilt binary / built from source with CMake, Release configuration]

What I did

I voxelized a 256³ grid as in the README.
On my setup, the measured time is about 4.7 ms for resolution 256, while the README mentions about 0.6 ms on a GTX 1050 Ti for the same resolution (excluding file I/O).

Because my card should be significantly faster than a 1050 Ti, I’m wondering if I am misunderstanding the benchmark conditions or missing some important build/runtime settings.

Questions

1. Are the numbers in the README pure kernel time (excluding file I/O and host↔device transfers), or something else?
2. Do you have recommended CMake / nvcc flags or CUDAARCHS settings for newer GPUs like RTX 40-series (e.g. sm_89)?
3. Are there any known issues or performance pitfalls when running this code with CUDA 12.x or on Ada / RTX 40-series GPUs?
4. Is there anything in the sample configuration (e.g. solid vs non-solid mode, specific test mesh) that I should be careful to match exactly?

If it helps, I can share more detailed logs, profiler output, or a minimal repro of how I am timing the kernel.

Thanks in advance for any hints!

[Zyl225]

CUDA VOXELIZER

CUDA Voxelizer v0.6 by Jeroen Baert
https://github.com/Forceflow/cuda_voxelizer - mail (at) jeroen-baert (dot) be

PROGRAM PARAMETERS

[Info] Filename: .\test_models\bunny.OBJ
[Info] Grid size: 256
[Info] Output format: magicavoxel file
[Info] Using CPU-based voxelization: No (default: No)
[Info] Using Solid Voxelization: No (default: No)

READ MESH

[I/O] Reading mesh from .\test_models\bunny.OBJ
[Mesh] Number of triangles: 5110
[Mesh] Number of vertices: 2557
[Mesh] Computing bbox

VOXELISATION SETUP

[Voxelization] Bounding Box: (-2.966815,0.033993,-2.477184)-(1.915775,4.916584,2.405406)
[Voxelization] Grid size: 256 256 256
[Voxelization] Triangles: 5110
[Voxelization] Unit length: x: 0.019073 y: 0.019073 z: 0.019073

CUDA INIT

[CUDA] CUDA device(s) found, picking best one
[CUDA] MapSMtoCores for SM 8.9 is undefined. Default to use 128 Cores/SM
MapSMtoArchName for SM 8.9 is undefined. Default to use Ampere
GPU Device 0: "Ampere" with compute capability 8.9
[CUDA] Best device: NVIDIA GeForce RTX 4060
[CUDA] Available device memory: 7099 of 8187 MB

TRIANGLES TO GPU TRANSFER

[Mesh] Allocating 179 KB of DEVICE memory for triangle data
[Mesh] Copy 5110 triangles to DEVICE memory
[Perf] Mesh transfer time to GPU: 0.4 ms
[Voxel Grid] Allocating 2 MB of DEVICE memory for Voxel Grid

GPU VOXELISATION

[Perf] Voxelization GPU time: 3.8 ms

FILE OUTPUT

[I/O] Writing data in vox format to .\test_models\bunny.OBJ_256.vox
[I/O] Writing to file: 0%...25%...50%...75%...100%

STATS

[Perf] Total runtime: 273.1 ms