What it does

This repository hosts pre-compiled TensorFlow wheels with CPU instruction-set optimizations (SSE4.1/4.2, AVX, AVX2, FMA) and various GPU/CUDA configurations. The builds target Ubuntu 16.04 LTS and cover TensorFlow versions 1.1 through 1.7, with both CPU and GPU variants across Python 2.7, 3.5, and 3.6.

The interesting bit

The README opens with the exact warning message you’ve probably seen and ignored: TensorFlow’s stock pip build leaving your CPU’s vector instructions idle like an unpaid intern. The project is essentially a build farm in repo form, with a combinatorial explosion of versions—CUDA 8/9/9.1, with/without MKL, MPI support, debug builds, even a lone macOS variant for 1.4.

Key highlights

• Compilation flags include -mavx, -mavx2, -msse4.1, -msse4.2, -mfma, and -mfpmath=both
• GPU wheels cover CUDA 8 through 9.1 with various compute-capability targets
• Versions 1.4.1+ add GCP, S3, and Hadoop support
• One-off builds exist: CPU debug wheels, MPI-enabled GPU builds, a single macOS CPU wheel
• Distributed via GitHub Releases with direct pip-installable URLs

Caveats

• Requires “relatively new Intel CPU” and recent Nvidia GPU; older hardware will simply not work
• Explicitly no Windows support (“we don’t have Windows machines ourselves”)
• Effectively frozen at TensorFlow 1.7; no updates for modern TF versions or Python 3.7+

Verdict

Grab a wheel if you’re stuck maintaining legacy TensorFlow 1.x code on Ubuntu 16.04 and want free CPU performance. Everyone else—especially anyone on modern TensorFlow—should compile their own or use conda; this is a historical artifact of the TF 1.x era.