Fun with triangles
The part of the hardware that created graphics in some form or another. To render things with them one often use some form of Graphical API. This pages focus on the way today's GPU's currently works, for past generations look at the GPU History page. Here is a list of the major manufacturers of GPU's.
Note: Working on writing this part so most of it incorrect :)
To confuse everyone including each other all the manufacturers have different names for the (almost) same things in a GPU. So i use my own name here and you can use the table below to find out what each vendor calls the each thing.
A GPU is made up of a custom processors (GPUCore) that execute programs in form of shaders. It can execute all types of shaders such as vertex shaders, pixel shaders or compute shaders.
Much of the work in a GPU is the same. The same shaders will need to run on all the vertices in a mesh and all the pixels in a triangle need to run the same pixel shader. The input for each one is different but the code to run is the same. To make use of that the GPUCore's are grouped with each other and each group is controlled by a GPUCoreMaster. All the cores in a group will run the same code and execute it in lock-step with each other. So if there is 32 cores in a group it will run 32 pixel shaders at the same time or vertex shaders at the same time. This is known as a SIMT architecture. If you only draw a cube with 8 vertices only 8 cores will do meaningful work in the group.
While executing there might be branch like if statements that send the cores down on different paths in the code. That is known as Divergence and the cores will keep running in lock-step and run the code inside the if statement. The cores that failed the statement will be masked out and throw away the result. When they exit the if statement the cores will converge and the masked out cores be activated again. Divergence lowers performance as some of the cores do wasteful work that are thrown away in the end.
A stall is when a core have to wait to run the next instruction. A common example is sampling a texture and waiting for it to return from memory. As the cores in a group run in lock-step they all have to wait for everyone to get the result back. This is solved with a form of threading and switching to another thread while waiting for the result of an operation. As the cores run in lock-step all threads in the group needs to be switched out at the same time. So each core runs a thread and all the threads running in a group in lock-step with each other is called a Wave. When a wave stalls the GPUCoreMaster can switch to one of the other wave's that are running on the group.
Thread - thread (NVIDIA) / work-item (AMD)
A thread is a single invocation of a program on the GPU. It can be a pixel shader or a vertex shader for example.
GPUCore - CUDA Core (NVIDIA) / Processing Element (AMD)
Wave - warp (NVIDIA) / wavefront (AMD)
Threads are executed in a group called a wave and all the threads in the wave execute the same instruction in lock-step.
GPUCoreMaster - Streaming Multiprocessor (NVIDIA) / Compute Unit (AMD)
Controls a group of GPUCores and run wave's on them.
GPU Programming - 2016
Visual Computing Systems - 2014
GPU Pipeline for Everyone - 2008
GPU versus CPU - 2008
A Closer look at GPUs - 2008
[Mobile] Graphics Hardware - 2007
3D Pipeline Of SM3/DX9 GPUs - 2006
Tiled hardware (speculations) - c0de517e.blogspot.ca
Intro to GPU Scalarization - 2018
GPU architecture resources
Optimizing for the RDNA architecture: presentation notes
GPU Optimization for GameDev
Unified Radeon™ GPU Profiler and Radeon™ Memory Visualizer usage with Radeon™ Developer Panel 2.1
Capturing GPU Work
Does subgroup/wave size matter?
Loads, Stores, Passes, and Advanced GPU Pipelines
GPU Captures: How we support placed and reserved resources
Nsight: The Most Important Ampere Tools In Your Utility Belt
Five years of GPU DB