Google DeepMind engineer shows five CUDA optimizations that speed up 3D Gaussian Splatting by 10.5x
janusch_patas · x · 2026-10-29
George Kopanas (Google DeepMind, co-author of the original 3DGS paper) published a deep dive presenting five self-contained CUDA optimizations that make forward/backward kernels up to 10.5x faster than the official Inria 3DGS implementation.
Key points:
- Baseline forward pass of 25.06 ms drops to 4.86 ms (5.2x) with literature optimizations, then to 2.39 ms (10.5x) via reduce-scatter, SH streaming, raster pruning, branchless prediction, and multirow (bf16) kernels.
- Even after combining every community optimization since release, the hardware ceiling isn't reached: another 1.78x speedup remains on the full forward-backward training step.
- Doubles as an end-to-end case study in optimizing a non-trivial GPU pipeline from first principles, aimed at graphics/ML practitioners.
Related event: DeepMind Engineer Speeds Up 3DGS Rendering 10.5x with CUDA Optimizations(3 posts)→
More from Infra
- Reve publishes technical report on training infra that scaled to 10K GPUs with 10x fewer chips — cantrell · 2026-10-29
- DeepSeek Flash fingerprint change hints at migration to a new GPU cluster — teortaxesTex · 2026-10-29
- Coding agents burn 95-99% cached tokens per benchmark task, token usage chart shows — AccBalanced · 2026-10-29
- OpenRouter's 10T Daily DeepSeek Tokens Are 95% Cache Hits; One GB300 Rack Could Serve It All — AccBalanced · 2026-10-29
- Kling 4.0 video model lands on Cloudflare AI Gateway at $0.157/sec for 720p — ritakozlov · 2026-10-29
- Google dodges TPU margin question, admits AI Mode queries are monetizing poorly — firstadopter · 2026-10-29