The speed of implementation

[cyw-3d]

Thanks for your torch implementation! It helps me a lot to reproduce the results of dreamfields.
When I compare your torch implementation with the official implementation written by JAX, I notice that your code runs much faster than the other. My question is where the acceleration comes from? Is it because pytorch is faster than JAX or do you use CUDA when you render?

[ashawkey]

The main acceleration comes from the CUDA ray marching introduced in instant-ngp to accelerate rendering.