About the training cost

[HUAFOR]

Could you kindly provide details on the type and number of GPUs used, as well as the training hours required for your model? Additionally, have you experimented with using smaller batch sizes during training?

Lastly, do you think it would be feasible to reproduce similar performance using eight 2080ti GPUs, each with around 12GB of VRAM, or would that configuration limit the model's efficiency?