DeepCPU: Serving RNN-based Deep Learning Models 10x Faster


Minjia Zhang, Samyam Rajbhandari, Wenhan Wang, and Yuxiong He, Microsoft AI and Research


Recurrent neural networks (RNNs) are an important class of deep learning (DL) models. Existing DL frameworks have unsatisfying performance for online serving: many RNN models suffer from long serving latency and high cost, preventing their deployment in production.

This work characterizes RNN performance and identifies low data reuse as a root cause. We develop novel techniques and an efficient search strategy to squeeze more data reuse out of this intrinsically challenging workload. We build DeepCPU, a fast serving library on CPUs, to integrate these optimizations for efficient RNN computation. Our evaluation on various RNN models shows that DeepCPU improves latency and efficiency by an order of magnitude on CPUs compared with existing DL frameworks such as TensorFlow. It also empowers CPUs to beat GPUs on RNN serving. In production services of Microsoft, DeepCPU transforms many models from non-shippable (due to latency SLA violation) to shippable (well-fitting latency requirements) and saves millions of dollars of infrastructure costs.

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.

@inproceedings {216077,
author = {Minjia Zhang and Samyam Rajbhandari and Wenhan Wang and Yuxiong He},
title = {{DeepCPU}: Serving {RNN-based} Deep Learning Models 10x Faster},
booktitle = {2018 USENIX Annual Technical Conference (USENIX ATC 18)},
year = {2018},
isbn = {978-1-939133-01-4},
address = {Boston, MA},
pages = {951--965},
url = {},
publisher = {USENIX Association},
month = jul,

Presentation Audio