Fact-checked by Grok 2 weeks ago

Accelerated Graphics Port

The Accelerated Graphics Port (AGP) is a high-speed, point-to-point interface standard developed by for connecting accelerators directly to a computer's , enabling faster data transfer between the and system compared to the shared bus. Introduced in to address the growing demands of and video applications, AGP allowed cards to access main more efficiently, reducing the need for expensive video while supporting pipelined operations and addressing for near-optimal bus utilization. Its initial specification (version 1.0) operated at a 66 MHz clock with 1× (266 MB/s) and 2× (533 MB/s) transfer modes using 3.3 V signaling, providing up to four times the of . Subsequent revisions expanded AGP's capabilities to keep pace with advancing graphics hardware. , released in 1998, introduced 4× mode (1.07 GB/s ) and 1.5 V signaling for better efficiency, along with features like Fast Write transactions and enhanced flow control to support deeper pipelining up to 256 requests. The final major update, 3.0 in 2002 (also known as AGP8X), doubled the peak to 2.1 GB/s in 8× mode through source-synchronous strobing and 0.8 V signaling, adding isochronous data transfer for low-latency applications and dynamic bus inversion to reduce noise, while maintaining with prior versions via a universal connector. AGP became the dominant graphics interface for consumer PCs from the late 1990s through the early 2000s, powering widespread adoption in and systems until it was progressively supplanted by (PCIe) starting around 2004. PCIe offered scalable, higher-bandwidth lanes (up to 4-8 GB/s initially) and multi-device support, rendering AGP obsolete by 2008 as manufacturers shifted entirely to the new standard. Despite its decline, AGP's innovations in dedicated graphics interconnects laid foundational groundwork for modern GPU architectures.

History

Development and Introduction

The Accelerated Graphics Port (AGP) was developed by in 1996 as a high-speed point-to-point designed specifically for attaching dedicated graphics cards directly to the , aiming to optimize performance for graphics-intensive applications. This approach addressed the limitations of shared bus architectures like , which constrained graphics data transfer rates during the emerging era of . The initial AGP specification, Revision 1.0, was released by on July 31, 1996, defining a 32-bit bus operating at 66 MHz that supported up to 533 MB/s of bandwidth when utilizing addressing for concurrent data and address transfers. led the development independently of the Special Interest Group but collaborated with leading graphics vendors, including and ATI, to ensure broad industry adoption through a royalty-free licensing model. The primary motivation stemmed from the mid-1990s boom in PC gaming and 3D graphics, where applications demanded faster access to system for textures and framebuffers to enable efficient rendering without excessive local video costs. AGP was publicly introduced in March 1997, with the first hardware implementations appearing later that year. , launched in August 1997, provided the initial motherboard support for AGP, integrating the interface into systems to accelerate 3D workloads. Concurrently, early AGP-compatible graphics cards emerged, such as NVIDIA's , released in August 1997, which leveraged the port's bandwidth for integrated 2D/3D acceleration and became a key driver in consumer adoption.

Adoption and Evolution

The Accelerated Graphics Port (AGP) saw rapid integration into consumer PC hardware following its specification release, with Intel's 440BX chipset providing early widespread support starting in early 1998, enabling 100 MHz front-side bus speeds alongside AGP 1x and 2x modes for enhanced graphics bandwidth over PCI. AMD followed with initial third-party chipset support via partners like VIA, but introduced native AGP 2x compatibility in its own AMD-750 chipset for Athlon processors in August 1999, broadening adoption across x86 platforms. By 1998, AGP had become the dominant interface for graphics accelerators in PCs, remaining the standard through 2004 as motherboard vendors like ASUS, Gigabyte, and MSI incorporated it into over 90% of mid-to-high-end boards, facilitating the shift from 2D to 3D-focused computing. AGP's adoption profoundly influenced graphics card development, with NVIDIA's , launched in October 1999 as the first dedicated GPU, leveraging AGP 4x to deliver hardware transform and lighting (T&L) capabilities that offloaded complex calculations from the CPU, supporting higher resolutions like 1024x768 in demanding titles. Similarly, ATI's Radeon 256 series, released in April 2000, utilized AGP to enable advanced pipelines with dual texture units and HyperZ bandwidth optimization, achieving up to 1 gigatexel/second fill rates for smoother gameplay at elevated settings. These cards exemplified AGP's role in enabling features like hardware T&L and , which were essential for rendering detailed environments in early games without bottlenecking system memory. The market evolved alongside AGP's versioning, transitioning from 2x modes (533 MB/s ) in chipsets to (1.07 GB/s) by 2000 with Intel's i815 and VIA KT133A, and culminating in 8x (2.13 GB/s) support in 2002 via Intel's 848 and AMD's 760MPX, correlating with 8's programmable shaders and titles like (). In , AGP configurations with cards yielded up to 30-50% frame rate gains over equivalents at high-quality settings with multiple lights and polygons, underscoring AGP's impact on real-time 3D gaming during its peak. Graphics vendors like and ATI adapted AGP protocols, incorporating fast writes and sideband addressing to optimize 3D pipelines for texture streaming and vertex processing, which became foundational for consumer 3D acceleration through 2004.

Decline and Obsolescence

The emergence of (PCIe) in 2004 marked the beginning of 's decline, as the new standard offered superior scalability through its lane-based architecture, allowing for configurable bandwidth allocation and support for multiple high-speed devices simultaneously. Unlike 's dedicated single-port design, PCIe enabled multi-GPU configurations such as NVIDIA's SLI and ATI's , which required parallel high-bandwidth connections that could not accommodate. NVIDIA accelerated the transition by adopting PCIe in its , launched in April 2004, while integrated PCIe support into its 9xx chipset series later that year, providing motherboard manufacturers with a unified interface for graphics and other peripherals. AGP's final major advancements occurred around 2003–2004, with the release of AGP 8x cards like NVIDIA's FX 5950 and ATI's Radeon 9800 XT, but production of high-end AGP graphics cards ceased by mid-2004 as manufacturers shifted focus to PCIe. By 2006, mainstream personal computers had fully phased out AGP slots in favor of PCIe, with NVIDIA's 7800 GS representing one of the last viable AGP options before the interface vanished from new motherboards. Several factors contributed to AGP's obsolescence, including bandwidth saturation at its 8x maximum of 2.1 GB/s, which proved insufficient for the escalating data demands of increasingly complex graphics workloads by the mid-2000s. Additionally, AGP's single-slot inherently lacked support for multi-GPU setups, limiting as gaming and professional applications increasingly relied on . Rising power and heat issues further exacerbated the problem, as high-performance GPUs exceeded AGP's standard power delivery limits of approximately 50 W—extended to 110 W only via the optional AGP Pro specification—necessitating auxiliary power connectors that PCIe integrated more efficiently through its 75 W per-slot capability and standardized auxiliary supplies. Today, persists primarily in legacy and niche applications, such as retro enthusiasts building period-accurate systems for early 2000s gaming, where compatible cards like the X800 remain sought after. In server repurposing, slots have been creatively adapted for non-graphics roles; for instance, a 2024 project demonstrated an -to- adapter on a motherboard, enabling the use of 66 MHz network cards and controllers for storage expansion, leveraging 's dedicated bridge for improved performance over standard . As of 2025, no active development or new standards for exist, solidifying its status as a legacy technology confined to hobbyist and archival contexts.

Design and Advantages

Improvements over PCI

The Accelerated Graphics Port (AGP) addressed key limitations of the bus for graphics-intensive applications by providing significantly higher dedicated to the graphics accelerator. While the standard bus offered a theoretical peak of 133 MB/s on a 32-bit interface at 33 MHz, AGP 1x achieved 266 MB/s at 66 MHz through optimized signaling, effectively doubling the available throughput without requiring the graphics card to compete with other peripherals for bus resources. Later versions scaled this further, with AGP 8x reaching 2.1 GB/s while maintaining , allowing sustained high-speed transfers for complex 3D workloads that could not support efficiently. A core architectural improvement was AGP's point-to-point design, which created a dedicated between the graphics card and the system's , in contrast to PCI's shared multi-device bus topology. This dedication eliminated arbitration delays and contention from other PCI devices, increasing data transfer throughput for graphics operations by up to an in pipelined scenarios. AGP's support for deep pipelining—allowing up to 256 outstanding requests—further minimized by enabling the graphics accelerator to issue multiple commands before receiving responses, a capability limited in PCI's simpler transaction model. AGP also introduced optimizations for memory access tailored to graphics needs, including optional sideband addressing via an 8-bit parallel port that separated address and data signals to increase efficiency during texture fetches. The Graphics Address Remapping Table (GART) enabled the graphics card to treat scattered system memory pages as a contiguous address space, facilitating direct access to main memory for textures and vertex data without the overhead of frequent local frame buffer swaps required on PCI. These features collectively accelerated texture mapping and geometry processing by shifting more data handling to cost-effective system RAM. In practical terms, these enhancements translated to faster frame buffer updates and improved performance. By isolating traffic, ensured consistent performance for emerging high-resolution and accelerated rendering applications, marking a pivotal advancement in consumer hardware integration.

Key Architectural Features

The Accelerated () employs a point-to-point bus optimized for , featuring a 32-bit multiplexed address/data (AD) bus operating at 66 MHz, which supports pipelined transactions to achieve high efficiency in data transfers between the and system memory. This structure includes an optional 8-bit address (SBA[7:0]) port that enables parallel address transmission, reducing compared to fully multiplexed in-band operations. The design's dedicated nature provides superior access for workloads over shared buses, though detailed contrasts are covered elsewhere. Central to AGP's graphics acceleration is the Graphics Address Remapping Table (GART), a memory-based table that maps scattered 4 KB pages of system RAM into a contiguous 32-bit space accessible by the (GPU) in AGP 1.0, with 64-bit support in later extensions, facilitating efficient and handling without excessive CPU intervention. AGP also incorporates a high-priority in and later for low-latency delivery of such as display refreshes through split transactions, with isochronous transfer added in AGP 3.0 for real-time applications. These features prioritize for graphics-specific operations, supporting prefetchable memory regions for patterns common in rendering. Signaling in operates in two primary modes to minimize overhead: in-band addressing uses the PIPE# signal to frame requests and commands on the AD bus, while sideband addressing leverages the SBA port for concurrent address delivery, which is optional for masters but required for targets and becomes mandatory in advanced configurations. Source-synchronous strobes (e.g., AD_STB[1:0] and SB_STB) synchronize data capture in higher transfer modes, enhancing reliability over the 66 MHz clock. AGP's electrical interface uses 3.3 V signaling in its initial version for compatibility with PCI environments, evolving to 1.5 V in subsequent revisions to support faster modes while maintaining backward compatibility via detection pins like TYPEDET#. The 3.0 specification further introduces 0.8 V signaling with dynamic bus inversion for 8x transfers, reducing power and electromagnetic interference in universal slots that tolerate both 1.5 V and legacy 3.3 V through dedicated power pins (VDDQ). This progression ensures robust, low-voltage operation tailored to graphics-intensive point-to-point connections.

Versions and Extensions

Standard AGP Versions

The Accelerated Graphics Port (AGP) evolved through three primary official versions, each building on the previous to increase bandwidth and efficiency for graphics data transfers between the CPU and graphics accelerator. AGP 1.0, released in 1996, established the foundational interface with support for 1x and 2x transfer modes operating at a 66 MHz clock rate and 3.3V signaling. In 1x mode, it provided a peak bandwidth of 266 MB/s, while 2x mode doubled this to 533 MB/s by transferring data on both rising and falling edges of the clock; these rates were calculated as effective throughput approximating (66 MHz × 32 bits × mode multiplier) / 8 bits per byte, enabling basic pipelining for up to 256 outstanding memory requests to reduce latency in graphics operations. AGP 2.0, introduced in May 1998, extended the specification by adding a mode, achieving up to 1.06 GB/s while maintaining with prior modes at either 3.3V or 1.5V signaling (with 4x requiring 1.5V for ). This version introduced the Fast Write , allowing direct CPU-to-graphics card transfers without intermediate buffering in system , which improved efficiency for and data uploads in 2x and 4x modes using PCI-like memory write commands with block flow control. Chipsets such as the 815 family supported AGP 2.0, providing 1x, 2x, and 4x modes at 1.5V/3.3V for enhanced graphics performance in mid-2000s systems. AGP 3.0, finalized in August , further doubled the maximum speed with an 8x mode delivering 2.1 GB/s at a 66 MHz clock, using a 1.5V with 0.8V peak-to-peak signaling to minimize noise and power consumption. for 8x was derived similarly as (66 MHz × 32 bits × 8) / 8 ≈ 2.1 GB/s, supporting up to 533 MT/s transfers, while incorporating improved error handling through features like Dynamic Bus Inversion (DBI) to reduce switching noise and isochronous error detection in the NISTAT for better reliability in high-throughput scenarios. This version was backed by chipsets like the 875P, which implemented 1x/4x/8x support for demanding graphics applications before the transition to .
VersionRelease YearMax ModePeak BandwidthKey Signaling VoltageNotable Features
1.019962x533 MB/s3.3VBasic pipelining (up to 256 requests)
2.019984x1.06 GB/s1.5V (for 4x)Fast Write protocol
3.020028x2.1 GB/s1.5V (0.8V swing)DBI, improved error codes

Official Extensions

The Accelerated Graphics Port (AGP) received official extensions from to address limitations in power delivery and memory addressing for high-end and professional graphics applications. One key enhancement was AGP Pro, introduced in 1998 and finalized in specifications released in 1999. This extension extended the standard AGP connector to support higher power requirements, enabling graphics cards with greater (TDP) suitable for demanding workloads. AGP Pro incorporated connectors, including an optional 8-pin and 4-pin configuration, to deliver up to 110 W total power—combining 25 W from the base slot with additional 12 V and 3.3 V rails from the extension. It also defined mechanical standards, such as I/O bracket designs for two- or three-slot cards and a minimum 200 linear feet per minute () for cooling, ensuring reliability in environments. was maintained, allowing standard cards to function in AGP Pro slots, though AGP Pro cards required the extended connector and were incompatible with legacy slots. These features targeted professional use cases like CAD and , where high-TDP GPUs needed stable power without relying solely on the slot. Another official extension involved 64-bit addressing support, integrated into the AGP 3.0 specification released in August 2002, with development beginning around 2001. This enhancement extended the effective address bus to 64 bits via the Graphics Address Remapping Table (GART), allowing access to memory beyond the 4 GB limit of 32-bit systems through features like 64-bit GART entries and the OVER4G bit in AGP status and command registers. The 64-bit mode doubled the addressable space for framebuffers without altering transfer speeds, relying on Dual Address Cycle (DAC) mechanisms from the PCI specification for compatibility. Implemented in select workstation and server chipsets, it facilitated extended memory mapping for large datasets in graphics-intensive server applications. For instance, AGP 3.0's 8x mode could leverage this addressing for improved performance in high-resolution rendering, though the core protocol remained unchanged from prior versions.

Unofficial Variations

Several manufacturers developed non-standard adaptations of the AGP interface to extend compatibility in systems lacking dedicated AGP slots, often integrating the protocol internally or via bridges to PCI or PCIe buses. These variations were not ratified by the official AGP specification and typically sacrificed full feature support for broader hardware integration. One common unofficial variation was the internal AGP interface, exemplified by Silicon Integrated Systems (SiS) chipsets such as the SiS 630 northbridge. This implementation used the AGP protocol internally to connect an integrated graphics core directly to the memory controller, eliminating the need for an external AGP slot and reducing motherboard complexity. Known as Ultra-AGP, it provided up to 2 GB/s memory bandwidth via an internal AGP 4x implementation while supporting features like pipelined transactions for onboard 3D acceleration, though it inherently limited upgradability since discrete AGP cards could not be added. A later iteration, Ultra-AGP II, increased maximum bandwidth to 3.2 GB/s in compatible SiS controllers. PCI-based AGP ports emerged through bridges like VIA Technologies' VT8601 Apollo ProMedia chipset, which incorporated an internal AGP-to-PCI bridge to enable AGP graphics operation over a PCI bus. This allowed AGP cards to function in systems without native AGP support by converting AGP signals to PCI transactions, supporting AGP 1.0 compliance with pipelined split-transaction bursts up to 533 MB/s. However, the bridge provided only partial feature implementation, such as basic GART support via a 16-entry TLB, but lacked higher-speed modes like AGP 4x or 8x. PCIe-based AGP ports relied on bridges to insert AGP cards into PCIe slots, with products like Albatron's ATOP released in 2005 using NVIDIA's BR02 chip to translate AGP 8x signals to PCIe x16. Motherboard vendors such as DFI and incorporated similar bridges in 2005-2006 boards (e.g., DFI's nForce4-based LanParty series with ULi southbridge extensions), allowing legacy AGP cards in transitioning PCIe systems. These s emulated AGP protocol over PCIe lanes but achieved only incomplete AGP 3.0 compliance, often capping at AGP 4x speeds. These unofficial variations shared key limitations, including reduced effective due to bridge overhead—e.g., VT8601's 533 MB/s versus AGP 8x's 2.1 GB/s—and compatibility issues with advanced features like Fast Writes or sideband addressing, leading to performance degradation of up to 20-30% in graphics-intensive tasks. Additionally, not all AGP cards functioned reliably, as often failed to handle signaling, resulting in or non-detection in some configurations.

Technical Protocol

Command and Request Mechanisms

The Accelerated Graphics Port (AGP) employs distinct mechanisms for initiating data transfers between the graphics accelerator and system memory, primarily through command codes and request packets transmitted via either in-band or addressing. These mechanisms support pipelined operations, allowing multiple outstanding requests to enhance throughput without blocking the bus. Commands and requests are framed using status signals ST[2:0], where the value "111" denotes the request phase on the AD bus. AGP command codes are 4-bit encodings (CCCC) that specify the type of operation, transmitted on the C/BE[3:0]# signals for in-band requests or as part of the address cycle. Common codes include 0000 for standard reads (transferring sequential Q-words starting at the specified address), 0100 for writes, 1100 for operations (ensuring all prior requests complete before subsequent ones), and 1010 for flush commands (to make pending writes visible to the system). In AGP 3.0, additional isochronous codes were introduced, such as 0011 for isochronous reads and 0111 for fenced isochronous writes, to support low-latency transfers. These codes are sent during the ST[2:0]="111" state to initiate transactions targeting system memory. No-ops are handled via patterns such as all 1s on SBA[7:0]. In-band AGP requests utilize the PIPE# signal to multiplex addressing and control information directly over the main 32-bit AD bus and 4-bit C/BE# lines, enabling compatibility with infrastructure while adding AGP-specific framing. When PIPE# is asserted low by the master (typically the graphics device), it enqueues one request per rising clock edge, with the (29 bits, aligned to Q-words) and 3-bit (LLL, specifying 1 to 8 Q-words or 8 to 64 bytes) placed on AD[31:3] and AD[2:0], respectively, alongside the command code on C/BE#. This method avoids the need for a separate address bus but limits throughput in higher modes due to contention with data transfers; it supports up to 256 outstanding requests, source-throttled by available slots. In-band addressing was deprecated in AGP 3.0, with PIPE# repurposed for other functions. Sideband AGP requests, in contrast, employ a dedicated 8-bit parallel bus SBA[7:0] to transmit and command information independently of the AD bus, reducing contention and enabling higher in 2x, , and 8x modes. The drives SBA[7:0] as outputs to the target (system logic), with requests structured in up to four sequential types per : Type 1 (lower bits [14:3] and LLL ), Type 2 (command code CCCC and mid- [23:15]), Type 3 (upper [35:24]), and optional Type 4 (extended for >64 GB). These are strobed via SB_STB (and SB_STB# in /8x modes) at source-synchronous rates, starting after a cycle where SBA[7:0] = FEh. Sideband addressing is optional in AGP 1.0 and 2.0 but mandatory in AGP 3.0, supporting pipelining without AD bus involvement. The request format incorporates a turn count mechanism to manage pipelining of multiple outstanding requests, specified via the RQ field in the AGPSTAT (values from 1 to 256), which indicates the target's maximum depth and prevents overflow. Data routing is embedded in the request, directing transfers from the graphics device to system memory (or vice versa) through the Graphics Address Remapping Table (GART) for aperture-mapped access, with priority sub-queues (high/low, in AGP 1.0/2.0) for read/write operations to optimize latency. For example, a read request routes data back to the GPU after completion, while writes push or command data directly to .

Response Handling

In the AGP protocol, responses are primarily managed by the device, which initiates signals to control , , and . The RBF# signal, asserted by the , indicates that the master's read is full, thereby stalling the return of low-priority read to prevent while allowing high-priority to proceed uninterrupted (in AGP 1.0/2.0). This control mechanism ensures efficient pipelining by requiring the master to have at least 16 bytes of buffering available (in 2x mode for transfers ≤16 bytes) before deasserting RBF# alongside GNT#. Similarly, the STOP# signal enables the target to abort a , such as during a disconnect or target-abort in Fast Write operations, prompting the master to deassert REQ# for at least two clock cycles unless it is the final . DEVSEL# is deasserted during native AGP s and is primarily used in compatibility modes to indicate device selection and transaction acceptance, with timing constraints that prevent master-aborts if decoded within the specified slow decode period; a turnaround is required if bus ownership changes following its use. Error handling in AGP emphasizes reliability for graphics workloads, where transient data reduces the need for extensive reporting. Parity errors trigger retries via underlying PCI protocols, as AGP does not independently detect or report PERR# during native transactions, instead relying on WBF# to block initiations if the write buffer is full. Fence commands provide pipeline synchronization by establishing boundaries in the master's access stream, ensuring that all preceding low-priority reads complete before subsequent writes, without consuming pipeline slots; high-priority requests may bypass these fences to maintain performance (in AGP 1.0/2.0). In AGP 3.0, fences further enforce ordering for isochronous transfers, guaranteeing that writes before a fence are globally visible before those after, often using shared memory buffers for additional synchronization. Completion of transactions occurs through data phases that match responses to original requests using identifiers like ST[2:0] encodings, which distinguish types such as low-priority reads (000) or high-priority reads (001) in 1.0/2.0, supporting split-transaction pipelining with depths up to 256 entries. This matching enables out-of-order processing while ensuring correct reassembly, with TRDY# and IRDY# handshakes at throttle points allowing either master or target to pause transfers by deasserting two clocks in advance. Burst transfers are completed in blocks of up to 256 bytes, aligned to 8-byte boundaries; in mode, this equates to 16 bytes per clock, with isochronous payloads scalable to 32, 64, 128, or 256 bytes without block-level stalling in AGP 3.0. The state machine governs these responses across defined phases to maintain protocol integrity. In the state, the bus is free with no outstanding requests, transmitting NOPs (all 1s at low voltage) to preserve signaling. The phase enqueues requests via PIPE# or Address (SBA) ports, latching GNT# and ST[2:0] for one clock. During the Data phase, actual transfers occur with signals, supporting 128-byte blocks in 8x mode for uninterrupted isochronous data. Turnaround phases insert dead clocks for bus ownership transitions or protocol shifts, such as from AGP to , ensuring clean signal propagation without contention.

Hardware Aspects

Connector Pinout

The Accelerated Graphics Port (AGP) connector is a specialized measuring approximately 1.5 inches (38 mm) in length, featuring 66 edge fingers per side for a total of 132 electrical contacts, designed to accommodate add-in cards up to 1.57 mm thick in a right-angle orientation on the . It incorporates keying to ensure voltage : 3.3V-only slots have a key between pins A22-A25 and B22-B25, 1.5V-only slots between A42-A45 and B42-B45, and universal slots have keying in the 3.3V position (A22-A25/B22-B25) but no key in the 1.5V position, allowing support for both 3.3V and 1.5V signaling environments through the TYPEDET# signal for automatic detection. This design maintains across AGP versions while preventing insertion of mismatched cards, with the universal variant enabling operation in all modes from 1x to 8x. Key signals on the AGP connector include the multiplexed 32-bit / bus AD[31:0], which handles both and transfers in a time-division manner; the 4-bit command/byte enable signals C/BE[3:0]#, used to specify types and byte enables; the optional 8-bit bus SBA[7:0], providing an independent for queuing additional requests to supplement the main bus; and the PIPE# signal, which selects between standard PCI-like addressing and pipelined AGP modes when asserted low. These signals are distributed across the connector's three rows (A, B, and C) to optimize , with strobe signals like AD_STB[1:0] and SB_STB ensuring source-synchronous timing for high-speed transfers in 2x and 4x modes. Power and ground pins are strategically placed for stable delivery and , with multiple +3.3V pins (labeled VCC3.3 or Vddq3.3) supplying the primary I/O voltage up to 1A per pin in 3.3V configurations, +1.5V pins (Vddq1.5) for low-voltage modes limited to 1A total, and +5V pins for with early peripherals though rarely used in core signaling. Ground (GND) pins are densely distributed, typically one per two signal pins, to provide low-impedance return paths and minimize , especially critical in 8x 3.0's 0.8V signaling environment. For Pro implementations, the connector extends on both ends with additional pins dedicated to auxiliary power: up to 7.6A on +3.3V and 9.2A on +12V (VCC12), enabling higher total power budgets up to 110W for demanding graphics cards without relying solely on the main slot. The pin assignments are organized into rows A (signals toward the card's leading edge), B (middle row), and C (trailing edge), with some pins reserved or no-connect (N/C) for future use. Below is a summarized table of representative assignments, highlighting key signals, power, and grounds; full specifications vary slightly by version (e.g., AGP 3.0 reuses some pins for DBI signals).
RowPinSignalDescription
A1Vddq3.3/Vddq1.5Primary I/O power (3.3V or 1.5V)
A2GNDGround
A3AD31Address/Data bus MSB
A11C/BE3#Command/Byte Enable MSB
A15SBA7Sideband Address MSB
A42PIPE#Pipelined mode select
A59AD0Address/Data bus LSB
A62CLK66 MHz clock
B1GNDGround
B2+12VAuxiliary power (AGP Pro extension)
B5INTA#Interrupt A
B11RBF#Read Buffer Full
B14SB_STBSideband strobe
C1+3.3V AuxAuxiliary power (AGP Pro)
C2GNDGround
C8GNT#Bus Grant
C22FRAME#Transaction frame (inverted in 3.0)
C35TRDY#Target Ready
C66GNDGround (trailing)

Compatibility Requirements

The Accelerated Graphics Port (AGP) requires specific hardware and software prerequisites to ensure full functionality, primarily centered on integration and voltage signaling compatibility. Motherboards must incorporate an AGP-compliant , such as Intel's i440LX introduced in , which supports the initial 1x mode at 66 MHz alongside interfaces. For higher speeds like 8x, later chipsets such as the Intel E7505 Hub are necessary, providing universal AGP 8x support with 0.8V signaling and backward compatibility to 1x through 4x modes. Additionally, a 66 MHz (FSB) is typically required for optimal performance in early implementations, as AGP operates synchronously with the system clock. The Graphics Address Remapping Table (GART) is a core requirement, implemented within the to map non-contiguous system into a contiguous for the , enabling efficient texture data . or must initialize the GART during (POST), configuring registers like ATTBASE for the table's base address (aligned to 4KB boundaries) and enabling via bits in the AGPCTRL register. Without proper support for GART allocation—often up to 256 MB by default—AGP cards may fall back to slower modes or fail to utilize system effectively. Compatibility between AGP cards and motherboards hinges on voltage matching to prevent electrical damage. Early AGP 1.0 and 2.0 implementations use 3.3V signaling for 1x and 2x modes, while 2.0's 4x and all of 3.0's modes (including 8x) employ 1.5V or 0.8V signaling. Universal slots, lacking a voltage key, support both via the TYPEDET# pin: grounded for 1.5V detection or open for 3.3V. Cards and slots negotiate the highest common mode during initialization, with hardware detecting capabilities via pins like MB_DET# and GC_DET# post-reset, followed by software programming of the AGPSTAT and AGPCMD registers to select rates (e.g., 010 binary for 8x). Strap pins on the connector, such as TYPEDET#, facilitate this automatic negotiation without user intervention. Software drivers are essential for mode detection and resource management. In and later, support integrates with via the , which handles GART setup and aperture allocation during driver initialization. For , the agpgart module provides chipset-specific backends (e.g., for or ), enabling GART and mode negotiation; it must be compiled into the or loaded as a module, with drivers recommending its use for non-supported chipsets to avoid conflicts. Common compatibility issues arise from mismatches, such as inserting a 3.3V-only into a 1.5V , which can overload and damage the due to excessive voltage; conversely, 1.5V in 3.3V slots are generally safe as they draw only the required lower voltage. Partial mode support occurs in mixed setups, where a 8x on a 2x downshifts to the lowest common rate, reducing without failure.

Power Consumption

The Accelerated Graphics Port (AGP) slot typically provides a base power delivery of 25 watts through its primary 3.3V and 5V rails in early implementations, with maximum currents of 6A on the 3.3V rail and 2A on the 5V rail, supporting low-to-mid-range graphics cards that draw around 20 watts in 1x modes. High-end cards operating at 8x speeds, such as those with advanced GPUs, can draw up to 75 watts, often requiring auxiliary power connectors to avoid exceeding slot limits. Early versions (1.0 and 2.0) rely on 3.3V signaling for 1x and 2x modes, with Vddq tolerances of 3.15V to 3.45V and maximum I/O currents up to 8A, enabling total slot power in the 25-50 watt range under full load. The 3.0 specification shifts to 1.5V signaling with Vddq at 1.425V to 1.575V and reduced maximum currents of 2A on the primary rail, significantly lowering power draw for the interface—estimated at about 50% reduction compared to 3.3V equivalents due to the lower voltage swing (0.8V peak-to-peak versus rail-to-rail in prior versions)—while capping signaling power at around 3 watts. This change prioritizes efficiency in higher-speed 4x and 8x operations without compromising bandwidth. Efficiency in AGP power usage is enhanced by features like sideband addressing, which separates address transmission from the main data bus (using dedicated SBA lines), reducing overall bus toggling and dynamic power dissipation compared to in-band addressing where addresses share the AD lines. In AGP 3.0, additional mechanisms such as Dynamic Bus Inversion (DBI) further minimize power by inverting data signals to limit simultaneous switching outputs, and logically inverted unasserted states keep quiescent power near zero. Heat dissipation remains a concern in dense layouts, where AGP slots' proximity to other components can exacerbate thermal buildup from sustained high draws. Power overdraw from incompatible or underpowered slots often leads to system instability, including artifacts, crashes, or failure to boot, particularly with demanding cards that exceed the 25-watt base limit. The AGP Pro extension addresses this by adding auxiliary 12V and 3.3V rails, supporting up to 110 watts total (e.g., 50-110W for Pro110 configurations), which was essential for power-hungry GPUs like the ATI Radeon 9700 series that required external supplementation beyond standard AGP capabilities.

References

  1. [1]
    [PDF] Accelerated Graphics Port Interface Specification
    Jul 31, 1996 · This is the Accelerated Graphics Port (AGP) Interface Specification, Revision 1.0, from Intel, dated July 31, 1996. Intel may have related ...
  2. [2]
    Intel runs graphics faster for cheaper - CNET
    Apr 2, 1996 · Intel's new Accelerated Graphics Port (AGP) technology for 3D acceleration ... 1996 and 1997 capable of delivering the best performance at ...Missing: announces | Show results with:announces
  3. [3]
    [PDF] A.G.P. 2.0 Specification
    May 4, 1998 · The A.G.P. 2.0 specification is the Accelerated Graphics Port Interface Specification, revision 2.0, from Intel, with a royalty-free license ...
  4. [4]
    [PDF] AGP V3.0 Interface Specification
    AGP3.0 is the new name for the AGP8X Interface Specification V1.0, introduced in August 2002.
  5. [5]
    What is AGP? The Differences Between 1x, 2x, 4x and 8x Explained | Lenovo US
    ### Summary of AGP from https://www.lenovo.com/us/en/glossary/agp/
  6. [6]
    The Graphics Bus Wars - IEEE Computer Society
    May 21, 2025 · The PCI was superseded by the accelerated graphics port (AGP) in 1997. [12] The AGP was developed to meet the increasing need for speed in 3D ...
  7. [7]
    Leading Graphics Chip Companies Announce AGP-Enabled Products
    The AGP specification, developed by Intel in cooperation with industry leaders, defines a high performance port that significantly improves the performance of ...Missing: 1996 | Show results with:1996
  8. [8]
    [PDF] Intel's 440LX Is First AGP Chip Set - CECS
    Aug 25, 1997 · But since the. 440LX works only with Pentium II, Pentium systems won't have access to AGP until other vendors deploy Pentium/AGP chip sets. ( ...Missing: chipset | Show results with:chipset
  9. [9]
    Famous Graphics Chips: Nvidia's RIVA 128 - IEEE Computer Society
    Aug 5, 2019 · By Jon Peddie It wasn't an April fool's joke in 1997 when Nvidia released the RIVA 128, based on the NV3 media accelerator.Missing: implementation | Show results with:implementation
  10. [10]
    AMD 750 Chipset Overview | PDF - Scribd
    AMD-750 TM. Chipset. Overview. Publication # 23016 Rev: A. Issue Date: August 1999 © 1999 Advanced Micro Devices, Inc. All rights reserved.
  11. [11]
    AGP Compatibility For Sticklers
    In 1996, Intel introduced AGP 1.0. AGP was a modified version of PCI designed to speed up transfers to video cards. It was followed by AGP 2.0 in 1998 and AGP ...Missing: adoption 1998-2004
  12. [12]
    Nvidia GeForce 256 celebrates its 25th birthday - Tom's Hardware
    Oct 11, 2024 · The original GeForce 256 used a 139 mm^2 chip packing 17 million transistors, fabricated on TSMC's 220nm process node. Contemporary CPUs of 1999 ...
  13. [13]
    ATI Radeon - Review 2000 - PCMag UK
    Rating 4.0 · Review by PCMag UKOct 17, 2000 · The 64MB Radeon card features ATI's Rage Theatre chip, which enables MPEG-1 and MPEG-2 video capture. The only shortfall with the Radeon is ...
  14. [14]
  15. [15]
    PCI Express* Architecture - Intel
    Introduced in 2004, PCIe* is managed by the PCI-SIG. PCIe* is capable of the following: Scalable, simultaneous, bi-directional transfers using one to 32 ...
  16. [16]
    Specifications - PCI-SIG
    PCI-SIG specifications define standards driving the industry-wide compatibility of peripheral component interconnects.PCI Express Specification · PCI Express 6.0 Specification · Ordering Information
  17. [17]
    PCI Express: An Overview - Ars Technica
    Jul 6, 2004 · The PCI bus debuted over a decade ago at 33MHz, with a 32-bit bus and a peak theoretical bandwidth of 132MB/s. This was pretty good for the time ...Missing: GeForce | Show results with:GeForce
  18. [18]
    Intel's 9xx Chipsets in the Flesh - Computex 2004 Early Bird Coverage
    May 31, 2004 · It may be that Intel rethought their PCI Express transition strategy and decided to include AGP support on their chipsets, we will find out more ...
  19. [19]
    Intel Doubles Graphics Bus Bandwidth With AGP 8x Technology
    Intel plans to launch two chipsets supporting AGP 8x for both dual processor and uni-processor workstations in the fourth quarter of 2002. A ...Missing: 1998-2004 | Show results with:1998-2004
  20. [20]
    Some questions regarding AGP power ... - AnandTech
    Oct 4, 2008 · The PCI-E Version 1.0 slot provides two voltages from the motherboard, 3.3V and 12V. You can have up to 3A on the 3.3V rail and up to 5.5A on ...Missing: delivery graphics<|separator|>
  21. [21]
    Modding ancient AGP slots is still oddly popular in 2024 to add ...
    Mar 10, 2024 · The AGP interface is newer and, thus, likely faster than the PCI interface on the motherboard, so using the first adapter is highly questionable ...Missing: multi- | Show results with:multi-
  22. [22]
    A Better Use For The AGP Slot, Decades Later - Hackaday
    Mar 3, 2024 · AGP is a superset of PCI clocked at 66 MHz, and usually benefiting from having its own exclusive bridge to the processor bus.
  23. [23]
    [PDF] AGP Speeds 3D Graphics; 6/17/96 - CECS
    Leveraging the PCI definition allowed Intel to develop the AGP specification more quickly; the very concept of AGP did not exist until six months ago. The ...
  24. [24]
    Intel Drives Accelerated Graphics Port (AGP) Interface Specification ...
    AGP Pro includes advanced capabilities such as high-performance single and multiple-image display, integrated video and 3-D functionality, and advanced realism.
  25. [25]
    Intel Announces Final Accelerated Graphics Port (AGP) Interface ...
    ORLANDO, Fla., July 21, 1998 – Delivering on a new graphics specification to extend the Accelerated Graphics Port (AGP) interface specification for advanced ...Missing: history | Show results with:history
  26. [26]
    [PDF] Intel(r) 815 Chipset Family
    — Support for Intel® Celeron® processors at 533A MHz and. ≥566 MHz (66 MHz ... The GMCH does not support any AGP/PCI access targeting the hub interface.
  27. [27]
    [PDF] 875P Chipset Datasheet - Intel
    The AMTT mechanism applies to the processor-AGP/PCI transactions as well and it assures the processor of a fair share of the AGP/PCI interface bandwidth.
  28. [28]
    [PDF] AGP Pro Specification - Motherboards.org
    AGP Pro extends the existing AGP connectors on both ends to deliver additional power on the 12 V and 3.3 V rails. The extension is illustrated in Figure 1.
  29. [29]
    [PDF] A.G.P. Pro Specification - ICminer
    The A.G.P. Pro definition includes an extended connector, thermal envelope, mechanical specifications for cards, I/O brackets, and motherboard layout ...
  30. [30]
    None
    Nothing is retrieved...<|control11|><|separator|>
  31. [31]
    [PDF] SiS630 Slot 1/Socket 370 2D/3D Ultra-AGPTM Single Chipset ...
    ... A.G.P. GART and Page Table Control Registers.......................... 110 ... SiS 630: Ready for Easy PC. 1.1. Function Block Reference Table. Bus ...
  32. [32]
    [PDF] VT8601 Apollo ProMedia
    Sep 8, 1999 · The chip also contains a built-in AGP bus-to-PCI bus bridge to allow simultaneous concurrent operations on each bus. Five levels.
  33. [33]
    Albatron Release AGP to PCI Express Adaptor? - PC Perspective
    New bridge card allows user to use AGP-8x card in a PCI-Express Mainboard Innovation Alert! Albatron's “ATOP” breathes new life ...
  34. [34]
    Albatron introduces AGP 8x-to-PCIe graphics card bridge - digitimes
    May 17, 2005 · Albatron has developed an AGP-to-PCI-Express (PCIe) bridge card called the ATOP, which allows users to use an AGP 8x VGA card on a PCIe ...
  35. [35]
    [PDF] Intel E7505 Chipset Memory Controller Hub (MCH)
    MCH uses Hub Interface 1.5 electricals on HI_A. HI_A also supports 64 bit upstream addressing via the hub interface extended address mechanism. 5.3. Hub ...
  36. [36]
    Chapter 12. Configuring AGP - NVIDIA
    AGP configuration uses the "NvAGP" option in X config, with values 0, 1, 2, or 3. Use the best driver for your chipset, and disable if stability issues arise.