What would happen if I half the batch_size and double the GPU number?

[ArchNew]

According to the paper, HRNet is trained on 4GPUs, which have larger memory.

If there are at most 8 GPUs available to me, but all have less than 12GB of memory. Can I double the GPU numbers and half the batch size to train the HRNet W48 with input resolution 384x288? Will it reach the same mAP?

One of my guides tells me, if the optimizer is SGD, my setting could reach the original results. But he doesn't know what would happen if I use the setting with Adam optimizer.