Intel Nervana Graph: A universal deep learning compiler
With the chaotic and rapidly evolving landscape around deep learning, we need deep learning-specific compilers to enable maximum performance in a wide variety of use cases on a wide variety of hardware platforms. Jason Knight offers an overview of the Intel Nervana Graph project, which was designed to solve this problem.
Talk Title | Intel Nervana Graph: A universal deep learning compiler |
Speakers | Jason Knight (Intel) |
Conference | Artificial Intelligence Conference |
Conf Tag | Put AI to Work |
Location | San Francisco, California |
Date | September 18-20, 2017 |
URL | Talk Page |
Slides | Talk Slides |
Video | |
With the chaotic and rapidly evolving landscape around deep learning, we need deep learning-specific compilers to enable maximum performance in a wide variety of use cases on a wide variety of hardware platforms. Jason Knight offers an overview of the Intel Nervana Graph project, which was designed to solve this problem. Intel Nervana establishes a hardware-independent intermediate representation (IR) for deep learning that all deep learning frameworks can target, which allows them to seamlessly and efficiently execute across present and future platforms with minimal effort. In addition to this IR, the project offers connectors to popular frameworks such as TensorFlow, Intel’s reference framework neon, and backends for compiling and executing this IR on CPUs, GPUs, and emerging deep learning accelerators.