deep learning model training time with different batch_size

Question

I was training resnet50 for classification with i9-13900k+4090, I found that when I set batch_size as 8, it took 3min+ per epoch, but when I set batch_size as 4, it only took 30s per epoch. I dont know why.Can you please tell me the possible reasons? Thanks in advance.

Dataset:350 images
Image Shape:512*360

I think might be something wrong with hardware, but have no idea.