Why is there a huge performance gap between torch implementation and tcnn?

[gjghs]

Hi.I tried both implementation setting on method instant-ngp

this is tcnn:
{
"experiment_name": "bear",
"method_name": "instant-ngp",
"checkpoint": "outputs/bear/instant-ngp/2024-01-11_010150/nerfstudio_models/step-000029999.ckpt",
"results": {
"psnr": 19.641199111938477,
"psnr_std": 1.3075436353683472,
"ssim": 0.5938013792037964,
"ssim_std": 0.07177797704935074,
"lpips": 0.3635721802711487,
"lpips_std": 0.07307428121566772,
"num_rays_per_sec": 299857.5,
"num_rays_per_sec_std": 74569.9140625,
"fps": 0.4087635576725006,
"fps_std": 0.10165316611528397
}
}
and this is torch:
{
"experiment_name": "bear",
"method_name": "instant-ngp",
"checkpoint": "outputs/bear/instant-ngp/2024-01-11_163252/nerfstudio_models/step-000029999.ckpt",
"results": {
"psnr": 18.415010452270508,
"psnr_std": 1.63877534866333,
"ssim": 0.4295531213283539,
"ssim_std": 0.08035850524902344,
"lpips": 0.5298488140106201,
"lpips_std": 0.09161611646413803,
"num_rays_per_sec": 52356.65234375,
"num_rays_per_sec_std": 4543.24609375,
"fps": 0.07137221097946167,
"fps_std": 0.006193319335579872
}
}

why is there a such huge performance gap.I thought the implementation would only affect the speed

The nerfacto method shows the same result.big gap between these implementations