Open Source GPU Compute Stack - Not Dancing the CUDA Dance

Explore the potential for creating an open-source, vendor-neutral GPU compute stack in this Linux Plumbers Conference talk. Delve into the current landscape of GPU acceleration technologies, including CUDA, ROCm, and OpenCL NEO. Examine the challenges and opportunities in developing a standardized alternative based on open-source technologies and development models. Learn about the proposed stack's components, including SYCL, OpenCL, Vulkan, and Mesa. Gain insights into execution environments, intermediate representations, and the development model for this ambitious project. Consider the implications for AI/ML, HPC, and open-source applications like TensorFlow. Understand the potential benefits and challenges of creating a unified compute stack that can work across multiple vendor GPUs.

Intro
Overview
Use cases
CUDA
Hat
OpenCL
Cycle
OpenMP
Vulcan
C standards
Compute stack overview
Execution environment
OpenCL execution environment
Intermediate Representations
OpenCL Stacks
OpenCL Vendor Stack
Rockem
Rockem runtime
Intel neo
Development model
Challenges
Proposed stack
Support costs
Cycles
Compilers
Smaller API
ABI API