Reproduce “A Scalable PyTorch Abstraction for Multi-GPU Gaussian Splatting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Matthew Cong, Francis Williams, Jonathan Swartz, Mark Harris, Sanja Fidler, Ken Museth
torch-dgx lets PyTorch models treat multiple GPUs as one aggregate device by distributing work inside supported operators through CUDA unified memory. For Gaussian splatting, it partitions Gaussian and image operations across GPUs, avoiding explicit cross-device communication in model code. It enables a reported 1-billion-splat aerial reconstruction that exceeds one GPU’s memory, demonstrating scale, not a matched comparison with prior work.
Key insights
- Efficient multi-GPU memory management relies on asynchronous unified memory pools, which significantly outperform synchronous allocation strategies by eliminating host synchronization.
- Abstracting multi-GPU functionality at the PyTorch device/operator level (`torch-dgx`) allows for transparent scaling of models without requiring explicit cross-device communication in the model's high-level code.
- Morton sorting provides only modest (5-10%) performance improvements for distributed Gaussian rasterization, indicating that 3D spatial locality does not perfectly translate to 2D screen-space locality during rendering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.