It is quite difficult to write code that extracts every ounce of performance out of a modern GPU. Doing so requires a deep understanding of the chip architecture and often lots of careful benchmarking and experimentation to see how best to map your workload onto the hardware. LLMs are exceptionally good at this type of software task, though, and these 3DGS speed ups show how much performance is sometimes sitting on the table!
I love too how the knowledge here is being shared openly and in a way will lead to a lot of rapid adoption (and widespread benefit for those who use this rendering stack). Thanks @gkopanas !
I love too how the knowledge here is being shared openly and in a way will lead to a lot of rapid adoption (and widespread benefit for those who use this rendering stack). Thanks @gkopanas !
Is it possible to make 3D Gaussian Splatting even faster?
Short answer YES!
Enjoy a 10.5x optimization on forward/backward cuda kernels over the published 3DGS ones!
grgkopanas.github.io/3dgs_cuda_blog/View quoted post on X
