Accelerated Graphics Port

from Wikipedia, the free encyclopedia
AGP slot

The Accelerated Graphics Port ( AGP ) is a connection standard on PC main boards ("Mainboard") for the direct connection of the graphics card with the chipset / northbridge . It is technically based on the PCI bus .


An AGP graphics card with an ATI chip

So-called “bursts” enable the data to be read quickly in this interrelated memory management. It also allows more efficient use of the framebuffer memory and thus also accelerates 2D graphics. Some RAID controllers use the AGP slot as a slot instead of PCI because of its higher transfer rate , although it was not actually developed for it.

There are a few motherboards with several independent AGP slots (each of which has to be assigned its own controller on the motherboard ), but normally there is only one AGP slot.


AGP was developed by Intel developed and was first in 1997 i440LX- chipset for the Pentium II - processor used. From 1998 AGP appeared on a large scale in commercially available PCs that were also equipped with chipsets from other manufacturers. The main board requires the AGP slot and the AGP bridge, which are usually located in the north bridge of the chipset. In this way, the graphics card is connected to the CPU and the RAM in the fastest way , both of which are essential for graphics performance.


The different versions of the AGP socket on the motherboard (slot bracket side on the left)

The first AGP version, AGP 1.0 , appeared in 1997 with the AGP 1x and AGP 2x modes. AGP 1x transmits 32 bits of data per clock with a clock frequency of 66 MHz, which is twice that of PCI , and achieves a theoretical transfer rate of 266 megabytes per second. AGP 2x transfers the data using the DDR method and achieves 533 MB / s with the same clock frequency. The signal voltage at AGP 1.0 is 3.3 volts. The fastest compatible cards are based on the ATI Radeon 9700 or Nvidia GeForce FX GPUs.

AGP 2.0 (1999) also offered the AGP 4x mode with a maximum transfer rate of 1066 MB / s with a reduced signal voltage of 1.5 volts. The increase in speed was achieved by now transmitting four data packets per cycle. Most graphics cards of this generation were fully backward compatible with the older standard and could be operated on motherboards with 3.3 V with AGP 2x. These models are called Universal AGP cards according to the AGP specification and have two notches in their connector strip.

In the new AGP 8x mode, AGP 3.0 (2002) achieved the highest AGP transfer rate of 2133 MB / s with a signal voltage of only 0.8 volts; instead of four, eight data packets were transmitted per cycle. The modes 1x and 2x no longer exist in this standard. Practically all graphics cards of this generation also support AGP 2.0 with 1.5 V and its slower modes ("Universal 1.5V AGP 3.0") for operation in older mainboards, some even support AGP 1.0 with 3.3 V ("Universal AGP 3.0" ) - this can be recognized by the two notches in the connection strip. Since the mechanical coding of AGP 3.0 models is identical to that of 2.0, all 3.0 cards tolerate installation in 2.0 systems, but non-universal cards do not start there. Older cards do not cause any damage in motherboards with the AGP 3.0 standard thanks to the protective circuit already provided in the standard, but only start in the universal models that are named.

AGP 3.5 (2004) did not introduce any new transfer modes, but instead focused on stabilizing implementation. Uniform standards for GART and Aperture (see AGP terms ) should bring a uniform AGP driver for all mainboards, which was also included in Windows XP SP2 in the form of the uagp35.sys file . The chipset-specific drivers used before repeatedly led to problems with AGP options such as FastWrites and faster transfer modes. Dual AGP mainboards for the interconnection of several graphics cards should also be possible, but this hardly had any effect due to the development of PCI Express and the multi-graphics card process based on it (e.g. Crossfire or SLI ).

In addition, there are cards based on the so-called AGP-Pro slot. Their contact strips are a little longer than the normal AGP cards. AGP Pro supplies cards with up to 50 W instead of the 25 watts of the normal slot. As an alternative to this, power-hungry graphics cards can be fed externally via a specially attached power supply (5 V / 12 V, Molex plug connector), which was the more frequently used solution due to the low distribution of the Pro slots. An AGP Pro slot also accepts normal AGP cards and was mostly found in the semi-professional workstation area.

AGP plug-in cards have a notch in the connector in which the corresponding bar in the slot on the motherboard must fit. With AGP 1.0 (AGP 1x / 2x) the bar sits further in the direction of the slot bracket, from AGP 2.0 it sits in the direction of the housing front wall. This is to prevent cards with 3.3 V signal voltage from being plugged into main boards with only 1.5 or 0.8 V signal voltage, as otherwise there is a risk of damage when commissioning (see next paragraph).


Universal AGP cards can be identified by the two notches in the plug contact

By assigning the bars to the signal voltages, inadvertent installation of incompatible graphics cards should be physically prevented. Nevertheless, there were both manufacturers of graphics cards and motherboards who installed an AGP universal connection, although their product was necessarily tied to an AGP version and thus voltage. A possible installation of incompatible cards leads to a high probability of damage to the graphics card and mainboard components.

Some TNT2 and SiS305-based graphics cards were manufactured that work with AGP 1.0, but with their misleading universal connector with two notches, they also fit in motherboards with a pure AGP 2.0 connector. However, these pure AGP 2.0 motherboards are incompatible with the voltage intended for version 1.0. Some of these main board types therefore have a protection circuit that prevents such 3.3 V cards from starting and causing damage. They usually indicate this acoustically or via LED.

There were motherboard models with Intel 845, Intel 850 and Nvidia nForce . These accidentally had a universal AGP port even though they do not support AGP 1.0. Here, too, the possibility of inserting a 3.3 V card usually caused voltage damage to both components. The use of these universal connections could be due to the actually universally compatible predecessor models of the chipsets. The Intel 815 chipset supports AGP 1.0 and 2.0 cards. It appeared two years before the i845, which cut support for version 1.0 and thus all 3.3 V cards.

As standard, AGP 3.0 cards have the same connection as 2.0 cards, but therefore all use the protective circuit provided so that they are not damaged when operated in 2.0 mainboards with their higher voltage. For compatibility see above paragraph on AGP 3.0.

AGP terms

Aperture size is an amount of main memory that the graphics chip can use to store textures via the AGP port . It is also referred to as AGP memory and should enable inexpensive graphics cards through lower amounts of local memory installed on the graphics card.

There are two modes of use, the names of which are often mixed up. In DiME mode ( di rect M emory E Xecute) the graphics chip can only use textures from the Aperture range to complete an image, its local memory is exclusively for rendering and frame buffer used. The Intel i740 chip used this mode to demonstrate the Aperture, although he also to Dimel mode ( di rect M emory E Xecute and L was able ocal Memory). This mode allows a sensible combination of local and AGP storage. For this, the most commonly used textures are kept local. Some PCI graphics cards were also able to do this before; With DiMEL, AGP achieved the advantage of faster AGP bandwidth to the main memory and the security that textures are not swapped out to the hard disk.

System memory refers to normal working memory outside the "aperture size", which is also available to AGP graphics cards like their predecessors. Probably the first chip to use this memory was the nVidia NV1 , which is exotic in many ways .

GART - The aperture size intended for exporting textures is only reserved in the main memory when required. It is previously available to the system for other purposes. If it is then used, the memory controller in the mainboard chipset will hardly find a contiguous memory section of this size. The data is therefore stored in a distributed (fragmented) manner. However, the graphics chip is provided with a contiguous memory. The translation is done by a GART ( Graphics Address Remapping Table ). This is similar to the memory management unit of a main processor .

Sideband addressing is the delayed transmission of addresses on the AGP. The graphics chip can request new data on these eight "sideband lines" while the current data is still being transferred on the 32 lines of the multiplexed address and data bus. The new data is then available immediately after the current one has been sent.

Fast Writes was only introduced with AGP 2.0 and is available for the AGP modes 2x, 4x and 8x. It was first implemented by the Nvidia Geforce 256 graphics cards. The graphics card and CPU can exchange data directly, instead of first writing them to the main memory and having the receiver read them again from there. This doubling of transfers without fast writes consequently effectively halves the RAM bandwidth. This comes into play as soon as the graphics card requests a corresponding amount of data. Fast-Writes fixes this and includes all data exchange, 2D and 3D. Often, however, the graphics card supports AGP modes that it can not utilize with its computing power . Especially in the entry-level segment, high AGP modes were implemented more because of their marketing effect. If the fast writes feature is missing here, it has just as little disadvantage as with all AGP mainboards, where the RAM bandwidth is about twice the AGP data rate (as with models with dual-channel DDR). In addition, game manufacturers usually paid attention to the most bandwidth-saving programming possible, so that the maximum utilization of the AGP and thus the need for fast writes came about less often. The challenge with fast writes implementation is a clean data signal. Therefore, Intel deactivated this function in the 815 chipset, as did the ATI Catalyst driver in systems diagnosed as unstable. It carries out a corresponding test after its installation. In general, it can help with graphics card driver problems to deactivate fast writes in the BIOS . The nouveau driver, which enables graphics cards from the manufacturer Nvidia to be operated under Linux , usually refuses to work when Fast Writes are activated. Fast writes are often possible by reducing the AGP mode where it was previously unstable, and can still help to increase the speed. However, this requires individual measurements using benchmarks.

The transition from AGP to PCI-Express

The standard PCI bus allowed a theoretical bandwidth of 133 MByte / s, cumulatively for all devices connected to this bus. In the early days (around 1995) this represented a gigantic range that could not be fully exploited. The first mainboards allowed transfer rates of max. 42 MByte / s with typical bandwidths to the main memory of almost 200 MByte / s.

With the increasing performance of CPUs and mainboards, these 133 MByte / s were largely exhausted, with graphics cards being the main consumers. This problem was solved with the creation of a dedicated AGP graphics bus and the construction of several PCI or PCI-X buses on server mainboards, but it led to a larger number of special slots on server mainboards with complex interactions due to non-point-to-point connections -Point connections between components. In addition, it remained at 133 Mbytes / s for the "rest" on normal mainboards, which increasingly individual components (e.g. copying from one IDE hard disk to another) were largely able to exploit.

With the development of PCI-Express, this development was straightened out again, there was a standard again, the bandwidth was almost doubled in the first step, the transmission switched from half-duplex to full-duplex, the bandwidth between devices was always available exclusively and never had to be shared .

PCI Bandwidth
(half duplex)
PCI Express Bandwidth
(full duplex)
PCI 32 bit / 33 MHz 0133 MB / s shared "Small consumers" PCIe 1.0 x1 0250 MB / s "Small
Consumer ", IEEE 1394b cards
PCI 32 bit / 66 MHz 0266 MB / s shared IEEE 1394b cards
PCI 64 bit / 66 MHz 0533 MB / s shared SCSI cards, ... PCIe 1.0 x4 1000 MB / s SCSI cards, ...
PCI-X 64 bit / 133 MHz 1066 MB / s PCIe 1.0 x8 2000 MB / s
AGP 8x 2133 Mbytes / s Graphics cards PCIe 1.0 x16 4000 Mbytes / s Graphics cards

With the AGP 8x version, this technology was developed to the end, as the speed cannot be increased any further due to timing problems with high clocks, which arise from the parallel data transmission, and the increasingly complicated circuit board design . The AGP feature of a fast point-to-point connection to the main memory had lost its importance, as graphics cards increasingly had their own RAM memory. In addition, AGP is only designed as an interface for graphics cards, but more and more other components, which until then were still connected via PCI or PCI-X , have higher requirements with regard to the data transfer rate. For this reason, among other reasons, the PCI-SIG designed the PCI Express standard as the successor to PCI and AGP, which offered an even higher data transfer rate than AGP. Windows 9x drivers were no longer supplied for the (after e.g. Socket 939 ) chipsets with DDR2-800 memory support . Hardly any new motherboards for AGP graphics cards have been introduced since mid-2006 . Graphics cards for an older 1-core processor AGP slot have become rare.

Web links

Commons : AGP  - collection of pictures, videos and audio files

Individual evidence

  1. AGP 3.0
  2. a b section "AGP Cards"
  3. Paragraph "AGP Motherboards"
  4. AGP 3.5
  5. Dave Dzatko, Tom Shanley, Mindshare Inc .: AGP system architecture . 2nd Edition. Addison-Wesley, 1999, ISBN 978-0-201-70069-5 , pp. 66 f .
  6. Archive link ( Memento from May 7, 2009 in the Internet Archive )
  7. Explanation of the GART translation
  8. УЗНАЙТЕ БОЛЬШЕ ( Memento from September 10, 2012 in the web archive ) Press kit for the first FastWrites implementation