Graphics processor

from Wikipedia, the free encyclopedia
Scheme with functional units of a modern graphics processor. For DIF see Video Display Controller .
NV20 GPU of an Nvidia GeForce 3

A graphics processor ( English graphics processing unit , or GPU for short ; this partially translates as a graphic unit and, more rarely, video unit or English video processing unit and visual processing unit , or VPU for short ) is a processor for computers and game consoles that is specialized and optimized for the calculation of graphics and smartphones. In addition, it outputs the calculated graphics to one or more displays. In the past, graphics cards had no computing abilities of their own and were purely output cards. From the mid-1990s onwards, 2D capabilities and later rudimentary 3D capabilities were integrated, the graphics processor was hard-wired or its programmability was limited to its firmware . Since the mid-2000s, the main processor (or CPU for short ) has been able to load programs onto the graphics card or into the graphics unit, which can thus be flexibly programmed to a limited extent.

GPUs can be found on the (sometimes of CPUs with integrated graphics unit APU , called on the) main board (Onboard as integrated graphics processor as well) on expansion cards ( plug-in card ). In the latter case, several GPUs are possible on one graphics card, or several graphics cards per PC. For notebooks there are external expansion boxes into which a graphics card can be plugged; A dedicated graphics unit can be built into docking stations . Almost all graphics processors for personal computers produced today come from AMD , Intel or Nvidia . The integration on plug-in cards, on the other hand, has been with other manufacturers for some time, with the exception of special and reference models.


A block diagram of the Radeon R300. Display interface (Eng. Screen interface) represents display controller + RAMDAC.

Display controller

To connect a screen to a computer - for example via MDA , CGA, etc. - you need a so-called screen adapter (analog network adapter). The chip on the (e.g.) ISA card is a comparatively simple video display controller . Any calculations for image synthesis are carried out on the CPU; the display controller simply packs this data stream into a corresponding signal (CGA, EGA, ...) for the screen. The card also contains a little memory, the so-called screen buffer. display buffer.


The RAMDAC is responsible for the conversion of digital data, which are present in the video memory / screen buffer, into an analog image signal.


The English called Graphics and Compute Array (short GCA ) or the "3D engine" (from the English , 3D engine ' borrowed can also) Graphic be designed calculations. The array also consists of the so-called shader processors , but also includes the geometry processor (see also geometry shader ).

Video compression / decompression

To reduce the amount of data of a video several are compression - algorithms have been developed. These describe extensive calculations which have to be carried out with the uncompressed data stream - which is already extensive in itself - in order to obtain the compressed data from it. To play a compressed video, appropriate calculations must be carried out on the compressed data stream.

These calculations can be carried out entirely or in part on the graphics card , the graphics processor or on another application-specific integrated circuit developed for this purpose.


Predecessors to graphics processors have been around since the early 1980s. At that time they only served as a link between the CPU and the screen output and were therefore called screen adapters (analog network adapters) video display controllers . They neither had the functionality, nor were they designed for independent calculations. Initially, they were primarily responsible for independent text and graphics output, thereby protecting the system bus . Some of them were later able to display sprites independently.

That changed in the mid-1980s with computers like the Commodore Amiga or the Atari ST . These already had blitting functions . In the x86 PC area, graphics processors with such additional functions emerged with the increasing spread of graphic interfaces, especially the Windows operating system . Blocks like the ET-4000 / W32 could process simple commands (e.g. "draw square") independently. Because of their main use under Windows, they were also called "Windows accelerators".

The first 3D accelerators hit the market in the mid-1990s . These graphics processors were able to perform some effects and triangle-based algorithms (such as Z-buffers , texture mapping ) and anti-aliasing . Such plug-in cards (such as the 3dfx Voodoo Graphics ) gave the computer games sector a boost in development. At the time, such applications were primarily limited by the processor.

The term GPU was first used intensively by Nvidia to market the Nvidia GeForce 256 series , which appeared in 1999 . This graphics card was the first (in the retail business) to be equipped with a T&L unit .

Due to their specialization in graphic calculations and concentration on massively parallelizable tasks, GPUs were and are usually clearly superior to CPUs in terms of their theoretical computing power. A CPU is designed for universal data processing; the individual CPU cores are also mostly optimized for fast processing of sequential tasks. The GPU, on the other hand, is characterized by a high degree of parallelization , since 3D calculations can be parallelized very well; therefore it specializes in 3D calculations, it is fast in calculations that use this functionality. The GPU still contains specialized units ("Fixed Function Units") for certain tasks (e.g. for texture filtering). A currently customary application program can i due to the lack of universality. A. Not running on a GPU. An algorithm that is limited to the capabilities of the GPU, but works very serially with relatively little data parallelism, cannot fully utilize the GPU. The relatively small caches in the GPU would lead to greater latencies in program execution, which, due to a lack of parallelization of the program, could not be compensated for by processing many tasks at the same time. The CPU is therefore faster for sequential tasks.

The performance advantage over CPUs for tasks that can be highly parallelized and the already existing SIMD properties make current GPUs interesting for scientific, graphic and / or data-intensive applications. This usage is known as GPGPU . The inclusion of the GPU has e.g. B. in the volunteer computing project Folding @ home led to an enormous increase in computing power. It was initially limited to chips from the manufacturer ATI / AMD, but in 2008 Nvidia GPUs from the GeForce 8 series were added. For graphics cards from Nvidia, CUDA exists as an API for using the GPU for calculations. This is now also used to perform physics calculations in games using PhysX . In the meantime there is the open programming platform OpenCL , with which programs for CPU and GPU can be developed equally. In addition, today's GPUs can calculate with double precision and not just with single precision .


The graphics processor takes on computationally intensive 2D and 3D computer graphics tasks and thereby relieves the main processor (CPU). The functions are controlled via software libraries such as DirectX or OpenGL . The released processor time can thus be used for other tasks.



Family (code name) Chip names Brand names (with renaming) installed in sales series (without renaming)
R100 R100, RV100, RV200, RS100, RS200 7xxx, 320-345
R200 R200, RV250, RV280, RS300 8xxx - 9250
R300 R300, R350, RV350, RV380, RS400, RS480 9500 - 9800, X300 - X600, X1050 - X1150, 200M ATI Radeon 9000 series , ATI Radeon X series , ATI Radeon X1000 series
R400 R420, R423, RV410, RS600, RS690, RS740 X700 - X850, X12xx, 2100 ATI Radeon X series, ATI Radeon X1000 series
R500 RV515, R520, RV530, RV560, RV570, R580 X1300 - X2300, HD2300 ATI Radeon X1000 series
R600 R600, RV610, RV630, RV620, RV635, RV670, RS780, RS880 HD2400 - HD4290 ATI-Radeon-HD-2000-series , ATI-Radeon-HD-3000-series
R700 RV770, RV730, RV710, RV740 HD4330 - HD5165, HD5xxV ATI Radeon HD 4000 series
Evergreen Cedar, Redwood, Juniper, Cypress, Palm ( AMD Wrestler / Ontario ), Sumo / Sumo2 ( AMD Llano ) HD5430 - HD5970, all HD6000s not listed under Northern Islands , HD7350 ATI Radeon HD 5000 series
Northern Islands Aruba ( AMD Trinity / Richland ), Barts, Turks, Caicos, Cayman HD6450, HD6570, HD6670, HD6790 - HD6990, HD64xxM, HD67xxM, HD69xxM, HD7450 - HD7670 AMD Radeon HD 6000 series
GCN 1.0 Southern Islands Cape Verde, Pitcairn, Tahiti, Oland, Hainan HD7750 - HD7970, R7 240, R7 250, R9 270, R9-280, R7 370, 520, 530 AMD-Radeon-HD-7000-series , AMD-Radeon-HD-8000-series , AMD-Radeon-R200-series , AMD-Radeon-500-series
GCN 1.1 Sea Islands Bonaire, Kabini , Kaveri and Godavari , Hawaii HD7790, R7 260, R9 290, R7 360, R9 390 AMD Radeon R200 series , AMD Radeon R300 series
GCN 1.2 Volcanic Islands Tonga, Fiji, Carrizo, Carrizo-L R9 285, R9 M295X, R9 380, R9 Fury AMD Radeon R300 series
GCN 4th generation Polaris RX 480, RX 470, RX 460, RX540 (X) - RX590 AMD Radeon 400 series , AMD Radeon 500 series
GCN 5th generation Vega RX Vega 56, RX Vega 64 (Liquid Cooled), Vega Frontier Edition, VII AMD Radeon Vega series
RDNA GPS RX 5700, RX 5700 XT AMD Radeon 5000 series
NV04 Fahrenheit Riva TNT, TNT2 Nvidia Riva
NV10 Celsius GeForce 256, GeForce 2, GeForce 4 MX Nvidia GeForce 256 series , Nvidia GeForce 2 series , Nvidia GeForce 4 series MX
NV20 Kelvin GeForce 3, GeForce 4 Ti Nvidia GeForce 2 series , Nvidia GeForce 4 series Ti
NV30 Rankine GeForce 5 / GeForce FX Nvidia GeForce FX series
NV40 Curie GeForce 6, GeForce 7 Nvidia GeForce 6 series , Nvidia GeForce 7 series
NV50 Tesla GeForce 8, GeForce 9, GeForce 100, GeForce 200, GeForce 300 Nvidia GeForce 8 series , 9 series , 100 series , 300 series
NVC0 Fermi GeForce 400, GeForce 500 Nvidia GeForce 400 series , Nvidia GeForce 500 series
NVE0 Kepler GeForce 600, GeForce 700, GeForce GTX Titan Nvidia GeForce 600 series , Nvidia GeForce 700 series
NV110 Maxwell GeForce 750, GeForce 900 Nvidia GeForce 900 series
NV130 Pascal GeForce GTX 1060, GeForce GTX 1070 (ti), GeForce GTX 1080 (ti), Titan X Nvidia Geforce 1000 series Nvidia GeForce 10 series
NV140 Volta GV100 Titan V, Quadro GV100
NV160 Turing GeForce RTX 2080 TI, GeForce RTX 2080 Super, GeForce RTX 2080, GeForce RTX 2070 Super, GeForce RTX 2070, GeForce RTX 2060 Super, GeForce RTX 2060 Nvidia Geforce 2000 series Nvidia GeForce 20 series

Power consumption

After the major manufacturers of CPUs began around the beginning of 2005 to reduce the power consumption of their products, in some cases very significantly, especially when the load was low, there was pressure on the manufacturers of graphics processors in this regard, but they have so far done the opposite: high-end graphics cards often convert more than 50 W into heat loss even without a load , although in this state there are practically no differences in performance compared to much simpler models or onboard graphics. At the end of 2007, AMD added efficient power-saving mechanisms to its desktop graphics cards for the first time with the ATI Radeon HD 3000 series . Nvidia developed the HybridPower process , which made it possible to switch off a high-end graphics card in 2D mode and switch to the economical onboard graphics chip, although a hybrid SLI-capable motherboard was a prerequisite for this. After a relatively short time, Nvidia said goodbye to this concept. In the meantime (2009) the GPUs of both manufacturers have mastered relatively efficient energy-saving mechanisms. ( See also: Green IT )


Market shares (number of units sold) of the GPU manufacturers (for desktop GPUs)


AMD (until 2006 ATI ), ARM Limited , Qualcomm , Intel , Nvidia , PowerVR

For years, Intel has been by far the market leader in graphics processors for PCs. The main reason is the large number of office computers, which are almost exclusively equipped with graphics processors integrated on the motherboard , which are part of the chipsets mainly supplied by Intel . AMD / ATI and Nvidia share the market in the field of card-based graphics solutions that are more suitable for PC gamers .


3dfx , 3DLabs , Cirrus Logic , Cyrix , D-Systems , Diamond Multimedia , Matrox , NeoMagic , Oak Technology , S3 Graphics , S3 Inc. , SiS , Trident Microsystems , Tseng Labs , XGI Technology Inc.

Due to the strong competition and the associated high development costs, most manufacturers have been bought up (3dfx, 3DLabs) or are concentrating on a niche market (Matrox, XGI).

Web links

Commons : Graphics processing unit  - album with pictures, videos and audio files

References and comments

  1. a b Mali-G52 / V52: ARM brings new GPU / VPU for AI smartphones - Golem , March 6, 2018
  2. The former graphics chip manufacturer 3DLabs used the abbreviation VPU to refer to the full programmability of fragment and vertex shaders in its products.
  3. RadeonFeature
  4. nouveau / CodeNames
  5. Power-guzzling graphics card: 78 boards in the test  ( page no longer accessible , search in web archivesInfo: The link was automatically marked as defective. Please check the link according to the instructions and then remove this notice. . Note: the GeForce 7600 GS, the most economical chip in the table, is said to still be 10 W in 2D and 20 to 27 W in 3D operation in a test by c't 04/2007.@1@ 2Template: Dead Link /