NervanaSystems/neonArchived
Neon
Neon is a deep learning framework and hardware-abstraction machine learning stack used for designing, training, and deploying neural network architectures. It functions as a graph-based computation engine that utilizes just-in-time kernel compilation to optimize machine code for tensors.
The platform decouples model definitions from execution kernels, allowing it to support multiple CPU and GPU backends. This architecture enables the distribution of computational workloads across parallelized hardware environments to increase processing speed and overall efficiency.
The system covers the full lifecycle of model development, from constructing custom neural network architectures to executing training routines and moving finalized models into production for real-time inference. Users can manage these tasks through API integration or by using configuration files to define parameters and select hardware backends.
Features
- Deep Learning Frameworks - Provides a complete system for the design, training, and deployment of neural network architectures.
- Computational Graphs - Implements a graph-based representation of mathematical operations to optimize tensor data flow and execution.
- Deep Network Training - Provides a hardware-accelerated stack for optimizing and training multi-layer neural architectures.
- Hardware Abstraction Layers - Provides an abstraction layer that routes tensor operations to multiple hardware backends like CPUs and GPUs.
- Just-In-Time Kernel Compilers - Utilizes just-in-time compilation to generate optimized hardware kernels for tensor operations at runtime.
- Machine Learning Platforms - Offers an integrated platform that decouples model definitions from execution kernels to support diverse hardware backends.
- Neural Network Construction - Provides APIs for designing and constructing deep learning architectures using layers and model abstractions.
- Neural Network Deployment - Facilitates moving finalized deep learning models into production environments for real-time inference.
- Neural Network Training Toolkits - Provides a software suite for the iterative development and optimization of deep learning models across parallelized hardware.
- Graph-Based Execution Engines - Implements an execution engine that optimizes mathematical operations by constructing directed acyclic graphs of tensors.
- Hardware Acceleration Backends - Provides hardware-specific compute backends to accelerate the execution of neural network computations.
- Hardware Backend Selection - Enables selection between different hardware processors, such as CPUs and GPUs, to balance execution speed.
- Custom Neural Architectures - Supports the implementation of tailored neural network structures built from low-level components and optimized kernels.
- Hardware Acceleration - Employs hardware acceleration techniques to offload mathematical computations to GPUs for faster model training.
- Tensor Memory Management - Implements specialized memory management for large multidimensional arrays to maximize throughput across CPU and GPU architectures.
- Multi-Backend Execution - Distributes computational workloads across parallelized hardware environments to increase training speed and efficiency.
- Deep Learning Frameworks - Python-based deep learning framework for performance.
- मशीन लर्निंग फ्रेमवर्क - Reference deep learning framework for Intel hardware.
- Data Science Tooling - High-performance deep learning framework.
