Facebook Opens Up Glow Compiler Back-End For Goya AI Accelerator

In the AI races for dedicated hardware accelerators, Habana Labs is off to an early lead when it comes to having a mainline, open-source kernel driver and now is also the first AI processor having a back-end implemented within Facebook's Glow AI open-source compiler.

Glow is the Facebook-developed compiler for neural network hardware accelerators designed for use by multiple machine learning frameworks. Glow is designed to be like an LLVM or Gallium of sorts for the AI space with interfacing between different ML frameworks and ultimately different hardware accelerators.

Glow, which is hosted under the PyTorch umbrella, is licensed under Apache 2.0. Their first experimental hardware back-end for the Glow compiler is Habana's Goya accelerator. Glow is developed by Facebook's AI team and this back-end was announced at this week's inaugural Glow Summit.

More details on code.fb.com.

Meanwhile with the in-development Linux 5.1 kernel is the initial Habana Labs kernel driver, the first for what ultimately the kernel developers are expecting to become the formation of an accelerator subsystem within the kernel tree.