GPU Accelerated Computing & Optimizations on Cross-Vendor Graphics Cards with Vulkan & Kompute
summary
Uses Vulkan SDK and Kompute to build a runnable GPU ML pipeline from scratch that executes on over 1000 GPUs across AMD Qualcomm NVIDIA, replaces the 500–2000 lines of raw Vulkan boilerplate with dozens of C++ lines, and demonstrates a working GLSL shader that multiplies tensors and returns correct results. Shows explicit device/queue selection and memory copies, implements an algorithmic example (logistic regression) and Android/Godot integrations, and proves that concurrency‑enabled GPU queues deliver 2x+ speedups on compatible hardware.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:15Objectives
- 2:36Why Parallel Processing
- 5:03Intuition
- 6:38CPU vs GPU Memory
- 9:35Grids Blocks Threads
- 10:47Leveraging Heterogeneity
- 12:30Vulkan SDK
- 14:08Vulkan Advantages
- 20:32Complexity Reduction
- 21:42Data
- 24:53Pipelines
- 26:13Sequential Program
- 29:56Compute Framework
- 32:42Compute Components
- 35:43Compute Manager
- 49:10Explicit Queues
related talks
18mgVulkan: Scalable GPU Pooling for Pixel-Grained Rendering in Ray Tracing
36mCUDA Kernels with C++
23mHello GPU
37mC++ for multi-accelerator, multi-vendor systems
25mVectorVisor: A Binary Translation Scheme for Throughput-Oriented GPU Acceleration
50mUnder the HAT: Empowering GPU Acceleration for Java
31mGPUs Unleashed! Make Your Games More Powerful With wasi-gfx
01h 32mHeterogeneous Programming in C++ today