Combining multiple GPU operations into a single optimized computation to reduce memory overhead and improve speed.