How do Graphics Cards Work? Exploring GPU Architecture

3-Second Blueprint (English)

1. A GPU has over 10,000 simple cores (10,752 CUDA cores on the GA102 chip) versus about 24 complex cores in a CPU — built for throughput, not latency.
2. Hierarchy: 7 graphics processing clusters with 12 streaming multiprocessors each; CUDA, tensor and ray-tracing cores do different jobs.
3. It excels at “embarrassingly parallel” work such as rendering and neural networks by running the same instruction over huge amounts of data (SIMD/SIMT).