Training deep learning models on ultrasound videos is like drinking from a firehose, most frames are redundant or noisy, yet drive high computational cost. We ask: can we learn just as well from less? We benchmark coreset construction, clustering, dataset distillation, and random sampling on large-scale echocardiography datasets. Clustering-based coresets, especially those using CNN embeddings,Wasserstein distance, and two-pass DBSCAN, match or surpass full-data training while reducing cost by up to 15×. They also rival dataset distillation in accuracy but are up to 30× faster. Surprisingly, random subsets sometimes outperform engineered coresets, with macro-F1 peaking at just 5% of the data. These results show that less can be more, offering a scalable path to efficient, fair training in medical video analysis.