TensorFlow 2.15 Is Faster, Brings New One-Shot Installer for NVIDIA CUDA Acceleration on Linux

The latest release of TensorFlow, version 2.15, is out now — and brings with it a much easier way to get started with CUDA-based accelerated machine learning on NVIDIA hardware under Linux.

“The tensorflow pip package has a new, optional installation method for Linux that installs necessary NVIDIA CUDA libraries through pip,” the TensorFlow team writes of its new software release. “As long as the NVIDIA driver is already installed on the system, you may now run pip install tensorflow[and-cuda] to install TensorFlow’s NVIDIA CUDA library dependencies in the Python environment. Aside from the NVIDIA driver, no other pre-existing NVIDIA CUDA packages are necessary.”

Built as the successor to Google Brain’s DistBelief. TensorFlow was first released in 2017 as a free and open source framework for machine learning and artificial intelligence workloads. In recent years the framework has been extended, now capable of running on everything from high performance computing clusters to microcontrollers — the latter using a special variant dubbed TensorFlow Lite for Microcontrollers and capable of performing on-device machine learning in resource-constrained environments.

Perhaps the biggest improvement to reach the platform in TensorFlow 2.15 is a way to drop the barrier to entry for new users working with TensorFlow on Linux: the ability to install TensorFlow with the libraries required to accelerate machine learning and artificial intelligence workloads on NVIDIA graphics processors in a single command. So long as the NVIDIA drivers themselves are installed, the TensorFlow team promises, accelerated TensorFlow will simply work.

Other improvements in the new release include performance improvements for oneDNN on Windows, an upgrade to CUDA 12.2 with performance improvements expected for NVIDIA Hopper-architecture graphics processors, and the move to Clang 17 as the default C++ compiler. TensorFlow 2.15 also makes tf.function types fully available, including tf.types.experimental.AtomicFunction — the fastest way, the team says, of performing TensorFlow computations in Python.

The new TensorFlow release is now available on GitHub, where it’s published under the permissive Apache 2.0 license.

Source link

%d bloggers like this: