Exploring the limits of Concurrency in ML Training on Google TPUs
arXiv:2011.03641
Abstract
Recent results in language understanding using neural networks have required training hardware of unprecedentedscale, with thousands of chips cooperating on a single training run. This paper presents techniques to scaleML models on the Google TPU Multipod, a mesh with 4096 TPU-v3 chips. We discuss model parallelism toovercome scaling limitations from the fixed batch size in data parallelism, communication/collective optimizations,distributed evaluation of training metrics, and host input processing scaling optimizations. These techniques aredemonstrated in both the TensorFlow and JAX programming frameworks. We also present performance resultsfrom the recent Google submission to the MLPerf-v0.7 benchmark contest, achieving record training times from16 to 28 seconds in four MLPerf models on the Google TPU-v3 Multipod machine.
References in corpus (8)
- Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
- Language Models are Few-Shot Learners
- Large Batch Training of Convolutional Networks
- Deep Learning Recommendation Model for Personalization and Recommendation Systems
- MLPerf Training Benchmark
- Addressing the Rare Word Problem in Neural Machine Translation
- Scale MLPerf-0.6 models on Google TPU-v3 Pods
- Automatic Cross-Replica Sharding of Weight Update in Data-Parallel Training