GPU Accelerated Computing on Cross-Vendor Graphics Cards with Vulkan Kompute

on YouTube ↗01h 02m

" summary

Shows concrete results: uses Vulkan and the Kompute framework to run a simple GPU‑accelerated ML kernel from scratch on AMD, Qualcomm, NVIDIA and mobile GPUs with only a handful of C++ lines, with working Android and Godot examples and a GitHub repo. Demonstrates async dispatch, explicit sequences and device selection, storage tensors for GPU‑only intermediates, and fixes for Raspberry Pi race conditions via memory barriers. Claims hardware concurrency yields 2x+ performance by submitting work to multiple GPU queues and verifies MoltenVK support on Apple.

AI-written summary · about

" chapters (19)

1. 0:00Introduction
2. 0:35Outline
3. 1:10Motivations
4. 3:00Intuition
5. 4:13Memory
6. 5:40Client Server
7. 7:06GPU Heterogeneity
8. 11:17Advantages and Disadvantages
9. 17:08Vulkan SDK
10. 24:10Hello World
11. 30:32Compute Architecture
12. 32:21Compute Example
13. 37:27Optimizations
14. 42:52Optimization Code
15. 47:21More Examples
16. 48:47Open Issues
17. 49:41Roadmap
18. 51:26Unified Memory Architecture
19. 51:39Vulkan SDKs

" related talks

• 01h 33mGPU programming with modern C++Michael S Wong · ACCU 2019
• 59mUsing Template Magic to Automatically Generate Hybrid CPU/GPU-CodeElmar Westphal · CppCon 2018
• 01h 32mHeterogeneous Programming in C++ todayMichael S Wong · ACCU 2018
• 01h 00mSYCL: Open Your Code to More Hardware ChoicesIgor Vorobtsov · C++ on Sea 2022
• 01h 01mA Modern C++ Programming Model for GPUs using Khronos SYCLGordon Brown · CppCon 2018
• 49mRapid Prototyping of Graphics Shaders in Modern C++Valentin Galea · CppCon 2018
• 01h 04mDesign for PerformanceFedor Pikus · CppCon 2018
• 39mThe CUDA C++ Developer ToolboxBryce Adelstein Lelbach · ACCU 2025