Tensorflow on Open Source GPUs

Explore the work done and remaining challenges in enabling Tensorflow deployment on open source Mesa drivers in this 43-minute conference talk from linux.conf.au 2020. Delve into the use of SYCL via LLVM and OpenCL, as well as efforts to enable OpenCL on a broader range of hardware. Learn about alternatives to the closed-source NVIDIA stack for AI/Machine Learning applications using GPUs, with a focus on open-source solutions. Gain insights into the ICON backend, execution methods, Vulkan shaders vs. kernels, the One API initiative, and the current state of open-source GPU compute for AI/ML frameworks.

Introduction
Outline
Tensorflow
ICON
Building
Backend
Execution
OpenCL
Vulcan
Shader vs Kernel
One API initiative
Sickle
Level Zero
Current State
Malcolm in the Middle
Clover OpenCL
Conclusion