Question about computation performance

[quan5609]

Hi, Thanks for the great and interesting work! I am looking forward to your open-source release. Currently, I have a question about the inference time performance reported in the main paper.

As mentioned in Table 1, your iNGP reproduction can reach 62 FPS on the Blender dataset. However, I am using iNGP source, and also in other papers such as https://hyperreel.github.io/ , iNGP is 10 times slower (5-6 FPS) for the same image resolution. Do you have any special technique for implementing grid-based NeRF or performance evaluation?