3D Gaussian Splatting CUDA overhaul claims 7.3× training-step speedup over Inria in nine-scene benchmark

The blog’s author reports a 7.30× speedup over Inria for the mean forward-and-backward training step across nine Mip-NeRF 360 scenes at the 30,000-iteration checkpoint on one Nvidia A100. The 10.5× figure highlighted in the accompanying post refers to the sum of per-stage timings on a single “room” test view, not the nine-scene mean. The author describes five hardware-level CUDA optimizations added after consolidating published techniques.