📄 中文摘要
该系统展示了深度学习可以在保持高准确度的同时显著加快运行速度。通过巧妙的技术，团队使每个GPU在进行更少复杂计算的情况下完成更多工作，从而实现快速学习而不损失模型质量。采用混合精度的方法，使得GPU能够进行更快的计算，同时保持结果一致。此外，团队还找到了一种使用非常大批量图像进行训练的方法，从而使整个集群能够扩展，并优化了机器之间的数据传输速度。最终，训练时间从过去的数十分钟缩短至仅几分钟，某些部分经过优化后相比于旧的设置实现了巨大的速度提升。
具有混合精度的高度可扩展深度学习训练系统：四分钟内训练ImageNet
📄 English Summary
Highly Scalable Deep Learning Training System with Mixed-Precision: TrainingImageNet in Four Minutes
The new system demonstrates that deep learning can run significantly faster while maintaining high accuracy. By employing clever techniques, the team enabled each GPU to perform more work with less complex calculations, allowing models to learn quickly without sacrificing quality. A mixed-precision approach was utilized, enabling GPUs to execute faster calculations while keeping results consistent. Additionally, the team discovered a method to train with a very large batch of images, allowing the entire cluster to scale up, and optimized communication between machines for faster data transfer. As a result, training time has been reduced from tens of minutes to just a few, with certain components optimized for huge speedups compared to older setups.