System Requirements¶
Hardware and software requirements for the 3DGS pipeline.
Minimum Requirements¶
Hardware¶
| Component | Minimum | Recommended | Notes |
|---|---|---|---|
| CPU | 4 cores | 8+ cores | Intel/AMD x86_64 |
| RAM | 16 GB | 32 GB+ | More for large datasets |
| GPU | NVIDIA GTX 1080 (8GB) | RTX 3090/4090/6000 Ada | CUDA required |
| VRAM | 8 GB | 24-48 GB | 48GB for 4K @ 5fps |
| Storage | 100 GB | 1 TB+ | SSD recommended |
| Network | 10 Mbps | 100+ Mbps | For downloads |
Software¶
| Software | Version | Purpose |
|---|---|---|
| OS | Ubuntu 22.04+ | Linux required |
| CUDA | 12.4 | GPU acceleration |
| Python | 3.10+ | Programming environment |
| Git | 2.0+ | Version control |
| FFmpeg | 4.0+ | Video processing |
Tested Configuration¶
Our Validated Setup:
Server: DNN19 (Mineno Laboratory)
CPU: Intel Xeon (multiple cores)
RAM: 128 GB
GPU: NVIDIA RTX 6000 Ada (48 GB VRAM)
CUDA: 12.4
OS: Ubuntu 22.04 LTS
Storage: 10 TB HDD
Results: - PSNR: 23.71 dB - Training time: ~18-30 min per date - Success rate: 100% (22/22 dates)
GPU Requirements¶
Why NVIDIA GPU?¶
CUDA acceleration is essential for: - COLMAP feature extraction (10-100x faster) - 3DGS training (required) - Rendering (significantly faster)
VRAM Requirements by Resolution¶
| Resolution | Frames (5fps) | VRAM Needed | Example GPU |
|---|---|---|---|
| 1080p | ~329 | 8 GB | GTX 1080 |
| 2K | ~329 | 12 GB | RTX 3060 |
| 4K | ~329 | 24 GB | RTX 3090 |
| 4K (our setup) | ~329 | 48 GB | RTX 6000 Ada |
VRAM Limitation
If you get OOM (Out of Memory) errors: - Reduce frame rate (5fps → 3fps) - Downscale resolution (4K → 2K) - Use fewer frames
Storage Requirements¶
Per Dataset¶
For one capture date (Condition 1, 5fps, 4K):
| Component | Size | Cumulative |
|---|---|---|
| Raw video | ~200 MB | 200 MB |
| Extracted frames | ~1.5 GB | 1.7 GB |
| COLMAP sparse | ~500 MB | 2.2 GB |
| 3DGS model | ~200 MB | 2.4 GB |
| Rendered images | ~2 GB | 4.4 GB |
| Total | ~5 GB | 5 GB |
For 22 dates: ~110 GB
Recommended: 500 GB - 1 TB for working space
Network Requirements¶
Download Sizes¶
| Item | Size | Time @ 100 Mbps |
|---|---|---|
| CUDA Toolkit | ~3 GB | ~4 min |
| COLMAP source | ~50 MB | ~5 sec |
| 3DGS repo | ~100 MB | ~10 sec |
| PyTorch + deps | ~2 GB | ~3 min |
| Total | ~5 GB | ~10 min |
Operating System¶
Supported¶
✅ Ubuntu 22.04 LTS (recommended, tested)
✅ Ubuntu 20.04 LTS (should work)
✅ Ubuntu 24.04 LTS (should work)
✅ Other Debian-based (may work)
Not Supported¶
❌ Windows (WSL2 may work but untested)
❌ macOS (no CUDA support)
❌ Other Linux distros (not tested)
Why Ubuntu?
- Best CUDA support
- Stable package ecosystem
- Wide community support
- Used in our lab
Software Dependencies¶
Required¶
# System packages
build-essential, cmake, git
libboost-all-dev, libeigen3-dev
libsuitesparse-dev, libfreeimage-dev
libgoogle-glog-dev, libgflags-dev
libglew-dev, qtbase5-dev
# Python packages
torch, torchvision (with CUDA)
numpy, pillow, tqdm
plyfile, opencv-python
# Tools
ffmpeg, conda/mamba
Optional¶
# For visualization
meshlab, cloudcompare
# For analysis
matplotlib, pandas, scipy
# For development
jupyter, ipython
Compatibility Check¶
Quick Test¶
Run these commands to verify:
# Check Ubuntu version
lsb_release -a
# Should show: Ubuntu 22.04
# Check GPU
nvidia-smi
# Should show: Your GPU model
# Check CUDA capability
nvidia-smi --query-gpu=compute_cap --format=csv
# Should show: ≥7.0 (Volta or newer)
# Check disk space
df -h
# Should show: ≥100 GB free
# Check RAM
free -h
# Should show: ≥16 GB
Performance Expectations¶
Training Time¶
| GPU | VRAM | Time (30K iter) | Cost |
|---|---|---|---|
| RTX 4090 | 24 GB | ~20 min | $1,600 |
| RTX 3090 | 24 GB | ~30 min | $1,000 |
| RTX 6000 Ada | 48 GB | ~18 min | $6,800 |
| A100 | 40 GB | ~15 min | $10,000 |
COLMAP Time¶
| Resolution | Frames | Time (feature) | Time (mapper) |
|---|---|---|---|
| 1080p | ~329 | ~5 min | ~10 min |
| 2K | ~329 | ~10 min | ~20 min |
| 4K | ~329 | ~20 min | ~40 min |
Upgrade Recommendations¶
If You Have 8 GB VRAM¶
Options: 1. Reduce resolution: 4K → 2K or 1080p 2. Reduce frame rate: 5fps → 3fps 3. Use fewer frames: Select best 200 frames
If You Have 16 GB RAM¶
Should work but: - Close other applications - Monitor memory usage - Consider 32 GB upgrade
If You Have Slow Storage¶
Impacts: - Frame extraction slower - COLMAP I/O bottleneck - Consider SSD upgrade
Cloud Alternatives¶
Don't have the hardware?
Google Colab¶
- GPU: Tesla T4 (16 GB) or A100 (40 GB)
- Cost: Free tier or $10/month Pro
- Limits: Session timeouts, storage limits
AWS EC2¶
- Instance: p3.2xlarge (V100 16GB)
- Cost: ~$3/hour
- Benefits: Full control, persistent storage
Paperspace¶
- GPU: Various options
- Cost: \(0.50-\)2/hour
- Benefits: Pre-configured ML environments
Cloud Considerations
- Data upload/download time
- Session management
- Cost for long training
- Storage fees
Verification Checklist¶
Before proceeding to installation:
- Ubuntu 22.04 LTS installed
- NVIDIA GPU with CUDA support (≥8 GB VRAM)
- At least 16 GB RAM (32 GB+ preferred)
- At least 100 GB free disk space
- Internet connection available
- Sudo/admin access to system
Next Steps¶
Hardware ready?
→ Quick Start Guide - Install everything in 30 minutes
Need help?
→ Troubleshooting - Common hardware issues
Hardware specifications based on DNN19 server configuration and testing with 22-date greenhouse dataset.