Adding gpus but without improvement of training speed

w86763777 / pytorch-ddpm

Unofficial PyTorch implementation of Denoising Diffusion Probabilistic Models

Do What The F*ck You Want To Public License

506 stars 62 forks source link

Adding gpus but without improvement of training speed #24

Open blablabLACK opened 4 months ago

blablabLACK commented 4 months ago

Hi, I am confused about the multi-GPU training speed.

Firstly, I trained the CIFAR10 dataset on my 2080 Max-Q (8G) GPU with default hyperparameters, and it took about 40 hours. Then I transferred the code to another computer equipped with 2 RTX-2080 Ti (12G) GPUs to train with the same parameters, but it still took 40 hours.

Why does this happen? Could someone give me some tips to improve the speed?

Thanks!