Memory & Speed Benchmark

We have compared Lightplane’s memory consumption and execution time of forward and backward passes with three baselines:

• PyTorch Autograd implementation

• PyTorch Autograd implementation with checkpointing

• NeRFAcc’s Instant-NGP

The following figure plots memory consumption and execution speed of each method as a function of the render size:

We can observe up to 1000x memory savings compared to the best baseline at similar (or even better) speed for both backward and forward passes.

Note

The Lightplane kernel benchmarks can be rerun using the renderer_speed_benchmark.py and splatter_speed_benchmark.py scripts under the tests directory.