Hardware

GPUs

A graphics processing unit (GPU) is circuitry optimised to perform the calculations required to display pixels. But GPUs are also used for other tasks that require many independent but similar calculations.

High resolution screens have millions of pixels, and each one’s colour needs to be calculated over and over again, ideally 60 times a second. The massively repetitive nature of the same calculations means this is a problem that can been speeded up by parallelism: performing calculations separately but simultaneously, and then combining them. GPUs have multiple processors built with specialised instruction sets optimised for the kind of calculations they need to perform. Common problems that GPUs need to solve rapidly include video decoding, and manipulation of three-dimensional geometry to produce two-dimensional views.

The current state of GPU technology has arisen partly because of the gaming industry. The computer’s main processors manage the running of the game, including maintaining the 3D model of the game world and the physics within it. This gets passed to the GPU that turns it into realistic 2D images fast enough to provide a smooth animation to the player. The GPU is performing the calculations needed to produce effects like perspective, occlusion, shadows, and distance-fading, as well as managing particles such as sparks or blades of grass.

Hardware acceleration

It’s increasingly common for a computer to have a GPU available to drive its monitor, so some tasks that might otherwise be handled by code running in the computer’s main processor can be handed over to the GPU. This is hardware acceleration: the task runs faster because the calculations are being done by the GPU’s hardware (that is, its specialised circuitry).

An example of this on the web is when animations are required (using CSS, or cascading stylesheets). Your web browser (running in the computer’s main processor) may do the calculations itself, or, if it has a GPU available, it may hand them over to benefit from hardware acceleration. You won’t necessarily know this is happening, but if it is you might notice that your machine doesn’t get bogged down, or that the animation is less choppy.

OpenGL and Shaders

The Open Graphics Library (OpenGL) is a public specification of the things GPUs can do. Most manufacturers make GPUs that support the OpenGL, which makes it feasible to access their capabilities from within any programming language that has bindings for them. Such bindings usually take the form of a library or plug-in that you can call from your own program.

Shaders are programs written specifically to run on GPUs.

OpenGL on the web: WebGL

Web pages can contain shaders written in the OpenGL Shading Language (GLSL ES). It’s a lot like the programming language C, but specialised for graphics and pixel-based calculations. The WebGL (Web Graphics Library) provides an application programming interface (API) in JavaScript for webpages. HTML’s canvas element can be manipulated in this way.

Programmers and the GPU

As a consequence of their parallelism, GPUs can be more efficient at certain types of non-graphical calculations than the processors of the machines they are connected to. Any calculation-based problem that can benefit from being run in parallel is potentially suitable for running on a GPU. Fields where such an approach is useful include machine learning (GPU manufacturer NVIDIA provides the CUDA platform for running Python frameworks like Tensorflow or Pytorch on GPUs) and cryptography (such as blockchain calculations, mining digital currencies, or brute-force attacks on encrypted data).