But Mummy I don't want to use CUDA - Open source GPU compute

A2 | Fri 25 Jan | 2:25 p.m.–3:10 p.m.


Presented by

  • Dave Airlie

    Dave is the Linux kernel graphics maintainer and has worked on improving the Linux graphics stack for more than 10 years. He works for Red Hat maintaining the graphics stack in all Red Hat distros. He is the initial creator of the virtio-gpu project, and the radv Vulkan driver for AMD GPUs.

Abstract

CUDA has become the defacto standard for GPU compute in most fields. AI and Machine Learning via projects like tensorflow are all targeting CUDA as their runtime. CUDA however as one major flaw, it's closed source and requires closed source drivers. This talk will explore the current state of non-CUDA compute stacks, concentrating on the OpenCL, SPIR-V and SYCL projects from Khronos, but also touching on Vulkan compute and other possibilities. It will discuss the some possible problems with AMD ROCm and Intel OpenCL drivers from a Linux platform point of view and try to scope out what a vendor-neutral Linux focused compute stack might look like that has maximal code sharing across vendors and doesn't require shipping various vendor forks of LLVM.