L01 GPU Fundamentals | Overview of GPU Hardware | Ep 2

This video is part of the playlist • hipFort Workshop . Click the link to view the entire playlist.
Accompanying notebook: github.com/pel...
The GPU's role is to bring in data, process it with instructions, and output the results. A GPU compute die shares many elements with a CPU. If we define a core as that which schedules instructions, then a Compute Unit in an AMD GPU is similar to a CPU core. The term Compute Unit comes from OpenCL and AMD, but in other terminology, it's known as a Streaming Multiprocessor (NVIDIA) or an XE Core (Intel). A GPU contains many compute units, and each one has a number of SIMD-like math units (typically 32-128), which perform computations on integers and floats in parallel with each other on vectors of data. These math units are called CUDA cores (NVIDIA), Shader cores (AMD), and XE vector engines (Intel). Each math unit is associated with a hardware thread that executes instructions issued by the compute unit. GPUs use the SIMT execution model, where hardware threads are grouped into teams.