PowerSGD

We study gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well or fail to achieve the target test accuracy.

We propose a new low-rank gradient compressor based on power iteration that can

1. compress gradients rapidly,
2. efficiently aggregate the compressed gradients using all-reduce, and
3. achieve test performance on par with SGD.

The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets.

Read feature

GitHub

Related Publications

PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization
Vogels, Thijs; Karinireddy, Sai Praneeth; Jaggi, Martin
2019-01-01	Advances In Neural Information Processing Systems 32 (NeurIPS 2019)