threestudio icon indicating copy to clipboard operation
threestudio copied to clipboard

Published 20 hours ago • verified publisher icon threestudio-project

Mixed precision (FP16) in NeRF code

Open claforte opened this issue 1 year ago • 1 comments

Hi, it seems like we're using FP32 all the time in nerf_volume_renderer. Since raymarching seems to be the primary bottleneck (both for speed and VRAM usage), it sounds like we could get a big speed-up and increase batch size by leveraging mixed-precision. Has anyone looked into it? The NerfAcc docs don't seem to mention mixed precision (support or lack thereof) but it looks like NerfStudio supports it... so it might be relatively straightforward? (I'll try to dig a bit deeper in case it's really easy.)

claforte avatar Jun 15 '23 18:06 claforte

Not easy but I might have found some folks who can get it done for us. ;-)

claforte avatar Jun 15 '23 20:06 claforte