This project is a deep learning framework designed for ease-of-use, extensibility, and high performance across various hardware. It targets researchers and developers needing a robust platform for training state-of-the-art models, offering significant speed advantages and a comprehensive model zoo.

How It Works

Neon is built with swappable hardware backends, allowing code to run on CPUs, GPUs, or Intel's specialized hardware. It leverages Intel's Math Kernel Library (MKL) for optimized CPU performance, providing significant speedups. The framework supports common deep learning layers and offers pre-trained models for various applications.

Quick Start & Requirements

• Install: git clone https://github.com/NervanaSystems/neon.git && cd neon && make (or pip install nervananeon for v2.4.0+). Activate virtual environment: . .venv/bin/activate.
• Prerequisites: Python 2 or 3. MKL is downloaded automatically. GPU backend selected by default if available; MKL backend selected by default on CPU.
• Resources: Setup involves cloning and making. Performance benchmarks show significant speedups over cuDNNv4.
• Docs: neon documentation

Highlighted Details

• Claims 2x speed of cuDNNv4.
• Offers a Model Zoo with pre-trained weights for models like VGG and Deep Residual Networks.
• Supports common layers including convolution, RNN, LSTM, GRU, and BatchNorm.
• Optimized for Intel CPUs via MKL for enhanced performance.

Maintenance & Community

• DISCONTINUATION OF PROJECT: Intel will no longer maintain or support this project. Patches are not accepted. Users are encouraged to fork.
• Community: neon-users Google group for questions and discussions.

Licensing & Compatibility

• License: Apache 2.0 License.
• Compatibility: Permissive license suitable for commercial use and closed-source linking.

Limitations & Caveats

The project is officially discontinued and no longer maintained or supported by Intel, meaning no future updates, bug fixes, or patches will be provided. Users needing ongoing support or development must fork the project. Manifest file format changed between v2.1.0 and v2.2.0, requiring recreation or updating for older manifests.