Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training

TLDR

Distributed training of deep models demands high communication bandwidth, limiting scalability and requiring costly high‑bandwidth networks, a problem exacerbated in federated learning on mobile devices with high latency and intermittent connectivity. This work demonstrates that 99.9 % of gradient exchange in distributed SGD is redundant and introduces Deep Gradient Compression (DGC) to dramatically reduce communication bandwidth. DGC preserves accuracy by applying momentum correction, local gradient clipping, momentum factor masking, and warm‑up training, and is evaluated on image classification, speech recognition, and language modeling across Cifar10, ImageNet, Penn Treebank, and Librispeech. Across these tasks, DGC achieves 270×–600× compression ratios, shrinking ResNet‑50 gradients from 97 MB to 0.35 MB and DeepSpeech from 488 MB to 0.74 MB, enabling large‑scale training on commodity 1 Gbps Ethernet and mobile devices without accuracy loss.

Abstract

Large-scale distributed training requires significant communication bandwidth for gradient exchange that limits the scalability of multi-node training, and requires expensive high-bandwidth network infrastructure. The situation gets even worse with distributed training on mobile devices (federated learning), which suffers from higher latency, lower throughput, and intermittent poor connections. In this paper, we find 99.9% of the gradient exchange in distributed SGD is redundant, and propose Deep Gradient Compression (DGC) to greatly reduce the communication bandwidth. To preserve accuracy during compression, DGC employs four methods: momentum correction, local gradient clipping, momentum factor masking, and warm-up training. We have applied Deep Gradient Compression to image classification, speech recognition, and language modeling with multiple datasets including Cifar10, ImageNet, Penn Treebank, and Librispeech Corpus. On these scenarios, Deep Gradient Compression achieves a gradient compression ratio from 270x to 600x without losing accuracy, cutting the gradient size of ResNet-50 from 97MB to 0.35MB, and for DeepSpeech from 488MB to 0.74MB. Deep gradient compression enables large-scale distributed training on inexpensive commodity 1Gbps Ethernet and facilitates distributed training on mobile.