Note: GPU support is available for Ubuntu and Windows with CUDA®-enabled cards.

TensorFlow GPU support requires an assortment of drivers and libraries. To simplify installation and avoid library conflicts, we recommend using a TensorFlow Docker image with GPU support (Linux only). This setup only requires the NVIDIA® GPU drivers.

These install instructions are for the latest release of TensorFlow. See the tested build configurations for CUDA® and cuDNN versions to use with older TensorFlow releases.

Pip package

See the pip install guide for available packages, systems requirements, and instructions. The TensorFlow package includes GPU support for CUDA®-enabled cards:

This guide covers GPU support and installation steps for the latest stable TensorFlow release.

Older versions of TensorFlow

For releases 1.15 and older, CPU and GPU packages are separate:

Hardware requirements

The following GPU-enabled devices are supported:

• NVIDIA® GPU card with CUDA® architectures 3.5, 3.7, 5.2, 6.0, 6.1, 7.0 and higher than 7.0. See the list of CUDA®-enabled GPU cards.
• On systems with NVIDIA® Ampere GPUs (CUDA architecture 8.0) or newer, kernels are JIT-compiled from PTX and TensorFlow can take over 30 minutes to start up. This overhead can be limited to the first start up by increasing the default JIT cache size with: '' (see JIT Caching for details).
• For GPUs with unsupported CUDA® architectures, or to avoid JIT compilation from PTX, or to use different versions of the NVIDIA® libraries, see the Linux build from source guide.
• Packages do not contain PTX code except for the latest supported CUDA® architecture; therefore, TensorFlow fails to load on older GPUs when is set. (See Application Compatibility for details.)

Note: The error message "Status: device kernel image is invalid" indicates that the TensorFlow package does not contain PTX for your architecture. You can enable compute capabilities by building TensorFlow from source.

Software requirements

The following NVIDIA® software must be installed on your system:

Linux setup

The instructions below are the easiest way to install the required NVIDIA software on Ubuntu. However, if building TensorFlow from source, manually install the software requirements listed above, and consider using a TensorFlow Docker image as a base.

Install CUPTI which ships with the CUDA® Toolkit. Append its installation directory to the environmental variable:

Install CUDA with apt

This section shows how to install CUDA® 10 (TensorFlow >= 1.13.0) on Ubuntu 16.04 and 18.04. These instructions may work for other Debian-based distros.

Caution:Secure Boot complicates installation of the NVIDIA driver and is beyond the scope of these instructions.

Ubuntu 18.04 (CUDA 10.1)

# Add NVIDIA package repositories # Install NVIDIA driver # Reboot. Check that GPUs are visible using the command: nvidia-smi # Install development and runtime libraries (~4GB) # Install TensorRT. Requires that libcudnn7 is installed above.

Ubuntu 16.04 (CUDA 10.1)

# Add NVIDIA package repositories # Add HTTPS support for apt-key # Install NVIDIA driver # Issue with driver install requires creating /usr/lib/nvidia # Reboot. Check that GPUs are visible using the command: nvidia-smi # Install development and runtime libraries (~4GB) # Install TensorRT. Requires that libcudnn7 is installed above.

Windows setup

See the hardware requirements and software requirements listed above. Read the CUDA® install guide for Windows.

Make sure the installed NVIDIA software packages match the versions listed above. In particular, TensorFlow will not load without the file. To use a different version, see the Windows build from source guide.

Add the CUDA®, CUPTI, and cuDNN installation directories to the environmental variable. For example, if the CUDA® Toolkit is installed to and cuDNN to , update your to match: