Training time?

[LyuJZ]

Hi, if I would like to reproduce the results with A6000 from Nvidia, how long will it take to train the model from scratch?

[sammychristen]

The training time largely depends on the hardware used.

On our hardware, pretraining takes around 22 hours and finetuning around 13 hours.

Best,
Sammy

[LyuJZ]

Thanks for your reply. So PointNet++ is only used to encode point clouds and not trained during reinforcement learning?

[sammychristen]

No, the PointNet++ is being finetuned as well during training.

[ychao-nvidia]

For reference, here is our benchmark for training time (in hours) on some GPU models.

	V100	A100	RTX 3090
Pretraining	48.8	22.2	21.5
Fintuning	30.9	13.5	12.7

Obviously, there could be many other hardware/system factors in play for time performance.

[Landy-929]

For reference, here is our benchmark for training time (in hours) on some GPU models.

V100 A100 RTX 3090
Pretraining 48.8 22.2 21.5
Fintuning 30.9 13.5 12.7
Obviously, there could be many other hardware/system factors in play for time performance.
May I ask that if I use the computer with 5070, will it run faster or normal? And if I did not use the Duplicate of use-ray in the instruction, it will take a long time, right?