Minimal requirements to get good results

[k0beLeenders]

I want to experiment a little bit with NeRF but when I use my gtx970 I get the resource exhaust error (understandable). I can get a gtx1070 but I'm not sure if this will suffice.

Any idea what would be the minimum requirements for training these models at a relatively fast speed? How much video memory do you need at least?

[merlinND]

Hi @KobeLeenders,

With the blender_config.txt provided in the repo (reproduced below), training a NeRF uses 8.9GB of GPU RAM. Usage is at 4.7GB at first, but after a few iterations it jumps to 8.9GB (maybe after doing the first preview re-rendering).

expname = blender_paper_lego
basedir = ./logs
datadir = ./data/nerf_synthetic/lego
dataset_type = blender

no_batching = True

use_viewdirs = True
white_bkgd = True
lrate_decay = 500

N_samples = 64
N_importance = 128
N_rand = 1024

[ghost]

The model is resolution-agnostic, so gpu memory shouldn't be an issue. You can reduce the number of rays processed per batch and the end result should stay the same, but it's slower. Look at tiny_nerf.ipynb, it is the chunk parameter in batchify(fn, chunk).