# TensorFlow: A system for large-scale machine learning

**Authors:** Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek G. Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke  
**arXiv:** [1605.08695](https://arxiv.org/abs/1605.08695)  
**Date:** 2016-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/1605.08695v1/

## Key points

- TensorFlow is a cutting-edge machine learning system designed for large-scale operations in diverse environments.
- It uses dataflow graphs to represent computation, shared state, and operations that modify the state.
- The unique architecture allows for efficient distribution across multiple machines and computational devices like CPUs, GPUs, and TPUs.
- TensorFlow excels in training and inference on deep neural networks with broad application support.
- Several Google services use TensorFlow in production environments, highlighting its reliability and performance.
- Released as an open-source project, TensorFlow has gained widespread adoption in the machine learning research community.

## Summary

, , , , 

TensorFlow is a cutting-edge machine learning system designed to operate at large scale and in diverse environments. It utilizes dataflow graphs to represent computation, shared state, and operations that modify that state. These graphs are efficiently distributed across multiple machines within a cluster and various computational devices such as multicore CPUs, general-purpose GPUs, and specialized ASICs like Tensor Processing Units (TPUs). This unique architecture provides developers with unparalleled flexibility to experiment with innovative optimizations and training algorithms, unlike traditional "parameter server" designs where shared state management is integrated into the system. One of TensorFlow's key strengths lies in its broad application support, particularly excelling in training and inference on deep neural networks. Several Google services leverage TensorFlow in production environments, underscoring its reliability and performance. Furthermore, TensorFlow has been released as an open-source project, gaining widespread adoption in the machine learning research community. Authored by a team including Martín Abadi, Paul Barham, Jeffrey Dean, Rajat Monga, and others, this groundbreaking paper delves into the intricacies of the TensorFlow dataflow model compared to existing systems. It showcases the remarkable performance that TensorFlow achieves across various real-world applications through detailed analysis and empirical evidence. With 18 pages and 9 figures providing comprehensive insights into the system's capabilities, this paper serves as a definitive resource for understanding how TensorFlow revolutionizes large-scale machine learning operations.