From Magic to Mechanics: Demystifying GPU Programming
summary
Explains why GPUs favor throughput over latency, gives a concrete hardware example (6,912 cores on an NVIDIA A100 computed from 108 SM × 64 cores), and walks through WebGPU’s 5-step initialization (canvas, adapter, device, context, configure). Shows runnable shader code and a working demo that renders a hardcoded red triangle, describes bind groups/buffers, warns that device–host transfers are the bottleneck, and argues WebGPU enables browser-side compute (inference) though native CUDA/Vulkan still leads for max performance.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 4:23Parallel programming basics
- 13:29The big picture of GPUs
- 16:18GPU architecture
- 27:03API landscape
- 30:25WebGPU fundamentals
- 39:27Pipelines and commands
- 44:57Debugging and conclusion
related talks
01h 06mFoundation of GPU Programming
38mWebGPU: Next-generation 3D Graphics on the Web
40mThe WebGPU Revolution: Full GPU Power on the Web and Beyond
34mWhat WebGPU means for Web platform?
36mWhy Are GPUs (Not) Fast - A Trip Through the Driver Stack
42mRendering performance from the ground up
24mArbitrary Computation on the GPU Using WebGL
32mGPU.JS - GPU Accelerated JavaScript