Question about training time in classic sr task

[lbc12345]

Hi, I'm interesting in this brilliant work, and I'm trying to reproduce the classic sr result. I'm using 8 2080ti GPUs and each GPU with batchsize 4. And after 2 days of training, my network only run about 200k iters per GPU. My questions are listed as follow:

1. Is batchsize 32 in your statement means total batchsize on 8 GPUs?
2. Is training iteration 500k means total iterations on all 8 GPUs? Or it is the iteration count on one single GPU?