Memory Occupancy

[develop-productivity]

Thank you for your outstanding work, and I wonder why use why to use the same vit - small architecture, dino would account for much more memory.
On my local computer, the batch size is 32 (1 GPU RTX 3090) takes up 17 G memory for dino-s/8. But it only take 12 G momory with batch size setted 128 for vit-s/16?
Thank you for your answer！