Going beyond eight million Gaussians on a 24 gigabyte VRAM GPU

[CVxTz]

Hello!

I am looking for advice or examples on fitting a Gaussian splat model with more than eight million Gaussians. Currently, the maximum amount of Gaussians I can fit on an ml.g5.8xlarge instance with 24 gigabytes of VRAM is around seven to eight million, and I want to try adding more to see if it improves visual quality.

The documentation mentions: "Automatic scaling to multiple GPUs in a single machine for large-scale reconstructions (source code only, packaging coming soon)."

I have a few questions regarding this:

• Multi-GPU Availability: Does this mean the multi-GPU feature is currently available in the main branch?
• OOM Prevention: Would utilizing multiple GPUs effectively help prevent Out of Memory errors during training when having more gaussians ?
• Activation: Is there a specific flag to activate multi-GPU training, or does the framework automatically detect and use all available GPUs?
• Memory Optimization: Are there any other recommended strategies, parameters to reduce memory usage during training on a single GPU?

Thanks in advance for your help!