The Accelerated Graphics Port (AGP) is a dedicated port which is targeted at 3D graphics applications. It allows for direct data transfer between the graphics card and the CPU and/or system memory. Although AGP is an extension of the PCI interface, it is completely (physically, logically and electrically) separate from the PCI bus. Therefore, activity of PCI peripherals won't affect the AGP card's performance.

AGP is based on the 66MHz PCI bus specifications but is further enhanced by :-

• allowing deeply pipelined memory read and write operations, thus masking memory access latencies
• demultiplexing (separating) address and data on the bus, thus allowing the bus to be more efficiently utilized
• allowing data transfers on both rising and falling edges of the signal, thus doubling the throughput of the bus (AGP2X)

The execution of 3D graphics applications requires the movement of textures, geometry data and commands (AGP requests) across the AGP bus to and from the graphics processor. To accomplish this, the AGP protocol offers several different modes.

Data Movement Paths and Corresponding AGP Modes (ã Intel)

Data written or read directly to or from the graphics controller is represented by the CPU-GC path. This path is typically used for CPU writes of geometry data and command lists to a memory-mapped region on the graphics card. According to the AGP 1.0 specification, data moved along this path must cross AGP using Frame Mode (PCI protocol).

Data written or read to or from main memory is typically texture data, geometry data or command lists stored in main memory. This path offers three mode choices: Frame Mode (GC-MMF), Pipelining Mode (GC-MMP) and Sideband Addressing Mode (GC-MMS).

Each mode offers a unique feature set and level of AGP performance. Frame Mode offers baseline performance using the PCI protocol. Pipelining Mode offers higher performance, and Sideband Addressing Mode offers the highest level of performance.

• CPU-GC Frame Mode — The AGP 1.0 specification allows only one master for AGP Pipelining and Sideband Addressing. Since this master is the graphics controller, only Frame Mode is available for moving data directly from the CPU to the graphics controller.
  
• GC-MMF AGP Frame Mode — This mode uses the PCI protocol to provide baseline performance for main memory reads and writes from the graphics controller. The higher bus frequency allows data transfers at 264 MB/s, or twice the throughput of shared 33-MHz PCI.
  
• GC-MMP AGP Pipelining Mode — AGP Pipelining offers higher performance by using the PIPE# signal of the AGP protocol to create multiple outstanding transactions, eliminating the bandwidth degradation due to round-trip time to main memory. Pipelined transfers are not coherent with the CPU's cache, so reads or writes are not delayed while the CPU caches are snooped to see if they contain the most recent copies of data being read or written. In a system with high AGP traffic, non-coherent transfers avoid heavily loading the system bus and allow data to be transferred more quickly to or from the graphics controller.
  
• GC-MMS Sideband Addressing Mode — This mode offers the highest level of AGP performance. In addition to allowing multiple outstanding transactions and non-coherent access to main memory, Sideband Addressing introduces a separate address/command bus, the Sideband Address Port (SBA). Because the SBA and data buses are not multiplexed, the graphic controller can use the SBA to make data requests without interrupting the data bus.

Pipelining and Sideband Addressing each offer two data rates—1X (66 MHz, 264 MB/s) and 2X (133 MHz, 528 MB/s). AGP Frame Mode can use only the 1X rate. At the cost of more high-speed board signals, the 2X rate almost doubles available bandwidth.

The topic of interest here would be the demultiplexing of the address and data buses or sidebanding. So, what's sidebanding actually?