but doesn't it break the assumption that it should ideally be able to reproduce your original results
To get CLIP to work properly we typically need large batch sizes. So the experiments in the original paper were quite heavy, and ran parallel across 8 GPUs.