[D] ViT's memory requirements, training time, and equivalent ResNet

My supervisor has asked me to try to create a table in which for each ViT model (ViT-s, ViT-b, ViT-l, and ideally Swin transformers), their estimated memory requirements (given some batch size), training time (based on arbitrary hardware) and on par ResNet model is specified.

I've been searching for quite a lot of time, and I absolutely can't find anything. Even the original ViT paper had no information in this regard. Do you think there's any way I can find …

hardware machinelearning memory requirements resnet searching swin table training transformers vit