JEST that trains based on entire batches is 13 times faster and 10 times more power efficient in comparison to typical training methods that use individual data points for training.

https://arxiv.org/pdf/2406.17711