# TVM: An Automated End-to-End Optimizing Compiler for Deep Learning

**Authors:** Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Meghan Cowan, Haichen Shen, Leyuan Wang, Yuwei Hu, Luis Ceze, Carlos Guestrin, Arvind Krishnamurthy  
**arXiv:** [1802.04799](https://arxiv.org/abs/1802.04799)  
**Date:** 2018-02-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.04799v3/

## Key points

- TVM is an open-source framework for machine learning optimization.
- It offers graph-level and operator-level optimizations, unlike current frameworks that rely on vendor-specific libraries.
- TVM addresses the challenge of deploying workloads to various hardware devices such as mobile phones, embedded devices, FPGAs, and ASICs.
- It incorporates high-level operator fusion, mapping to hardware primitives, and memory latency hiding techniques for deep learning optimization challenges.
- TVM uses a learning-based cost modeling method to automate low-level program optimization based on hardware characteristics.
- Experimental results show competitive performance across different hardware back-ends including CPUs, mobile GPUs, and server-class GPUs.
- TVM can target new accelerator back-ends like FPGA-based deep learning accelerators.
- The team behind TVM includes authors Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Meghan Cowan, Haichen Shen, Leyuan Wang, Yuwei Hu,Luis Ceze,C arlos Guestrin,and Arvind Krishnamurthy.
- Their work on TVM represents a significant advancement in automated optimization for deep learning compilers.

## Summary

TVM is an  for , providing a solution to the growing demand for bringing machine learning to a wide range of hardware devices. Unlike current frameworks that rely on vendor-specific operator libraries and are optimized for specific server-class GPUs, TVM offers  by exposing graph-level and operator-level optimizations. It addresses the challenge of deploying workloads to new platforms such as mobile phones, embedded devices, and accelerators like FPGAs and ASICs. Specifically designed for deep learning optimization challenges, TVM incorporates high-level operator fusion, mapping to arbitrary hardware primitives, and memory latency hiding techniques. It also utilizes a novel learning-based cost modeling method to automate the optimization of low-level programs based on hardware characteristics. This allows for rapid exploration of code optimizations. Experimental results have shown that TVM delivers competitive performance across various hardware back-ends including low-power CPUs, mobile GPUs, and server-class GPUs. It has also demonstrated its versatility by targeting new accelerator back-ends such as FPGA-based generic deep learning accelerators. The system is open-sourced and already in production use within several major companies. The team behind TVM includes authors Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Meghan Cowan, Haichen Shen, Leyuan Wang, Yuwei Hu,Luis Ceze,C arlos Guestrin,and Arvind Krishnamurthy. Their work on TVM represents a significant advancement in automated optimization for deep learning compilers.