A graphics processing unit (GPU) is a processor designed to handle graphics and other highly parallel workloads. GPUs are best known for drawing images, video, and 3D scenes, but modern GPUs are also used for scientific computing, machine learning, video encoding, simulation, data analysis, and other tasks where the same operation can be applied to many pieces of data at once.
A GPU may be built into a processor, fitted as a separate graphics card, included in a console, or integrated into a phone, tablet, workstation, server, or embedded device.
Purpose
The central processing unit is usually built for flexible general work, fast branching, operating system tasks, and low-latency control. A GPU is built around wide parallel throughput. It can process large numbers of pixels, vertices, shader operations, matrix calculations, or compute jobs at the same time.
This design is useful for graphics because an image contains many pixels that can be processed in parallel. It is also useful outside graphics when an application can divide a problem into thousands or millions of similar operations.
Integrated and Discrete GPUs
An integrated GPU shares a package or system memory with the main processor. Integrated graphics are common in laptops, office PCs, handheld devices, and low-power systems. They are efficient and compact, but usually share power and memory bandwidth with the CPU.
A discrete GPU is a separate chip, often mounted on a graphics card with its own video memory. Discrete GPUs are common in gaming PCs, workstations, render nodes, and accelerated servers. They can draw more power and require more cooling, but they usually provide stronger graphics and compute performance.
Architecture
GPU designs vary by manufacturer and generation, but most modern GPUs contain many execution units arranged for parallel work. A graphics pipeline may include stages for geometry, rasterisation, shading, texture sampling, blending, and output to a frame buffer.
Video memory stores frame buffers, textures, geometry, shader data, and other resources. Higher memory bandwidth is important because graphics and compute workloads often move large amounts of data.
Modern GPUs may also include:
- Ray tracing hardware for lighting, reflections, and acceleration structures.
- Video encode and decode blocks.
- Display controllers.
- AI or matrix acceleration units.
- Cache systems tuned for graphics and compute access patterns.
Software Interfaces
Applications usually reach the GPU through an application programming interface. Major graphics and compute interfaces include:
- DirectX, used mainly on Microsoft platforms and Xbox.
- Vulkan, a cross-platform graphics and compute API from Khronos.
- OpenGL, an older but widely deployed graphics API.
- Metal, used on Apple platforms.
- CUDA, NVIDIA's parallel computing platform and programming model.
- OpenCL, a cross-platform compute framework.
The API does not replace the GPU driver. The driver translates application requests into commands that the hardware can execute.
Uses
GPUs are used in several areas:
- Games: drawing 2D and 3D scenes, shaders, lighting, post-processing, and high frame rates.
- Video and media: playback, encoding, decoding, colour conversion, and editing previews.
- 3D creation: modelling, rendering, animation, visual effects, CAD, and digital content work.
- Scientific computing: simulations, numerical models, molecular work, weather models, and data processing.
- Machine learning: training and running models that use large matrix operations.
- Servers and cloud systems: accelerated compute, virtual desktops, media processing, and AI workloads.
Manufacturers
NVIDIA and AMD are major suppliers of discrete GPUs for gaming, workstations, and servers. Intel supplies integrated graphics in many processors and also sells Intel Arc discrete GPUs. Apple, Qualcomm, Arm partners, and other companies supply GPUs in phones, tablets, laptops, consoles, and specialised systems.
See Also
References
Discussion log
Use comments for sourcing notes, corrections, and disputed details.
No comments yet.