Overview
Explore the landscape of open-source GPU computing alternatives to CUDA in this 43-minute conference talk from linux.conf.au. Delve into the current state of non-CUDA compute stacks, focusing on OpenCL, SPIR-V, and SYCL projects from Khronos, as well as Vulkan compute and other possibilities. Examine potential issues with AMD ROCm and Intel OpenCL drivers from a Linux platform perspective, and consider the prospects of a vendor-neutral, Linux-focused compute stack that maximizes code sharing across vendors without requiring various vendor forks of LLVM. Gain insights into the anatomy of GPU compute stacks, including building, application, execution, and OpenCL specifics. Compare vendor stacks like CUDA, ROCm, and NEO, and understand the distribution challenges faced by open-source alternatives. Conclude with a proposed stack and reflections on the future of open-source GPU computing.
Syllabus
Intro
Use Cases
API - CUDA
API - HIP
API- OpenCL
API - Others
Future APIs - C++2?
Stack Anatomy (Building)
Stack Anatomy (Application)
Stack Anatomy (Execution)
Stack Anatomy (OpenCL)
OpenCL Stacks
Vendor Stacks - CUDA
Vendor Stacks - ROCm
Vendor Stacks - NEO
Distribution Challenges
Proposed Stack
Conclusions??
Taught by
linux.conf.au