Comparison between the PCI and AGP bus

Comparison between the Peripheral Component Interconnect (PCI) and the Accelerated Graphics Port (AGP)

AGP will deliver a peak bandwidth that is 4 times higher than the PCI bus using pipelining, sideband addressing, and more data transfers per clock.
It will also enable graphic cards to execute texture maps directly from system memory instead of forcing it to pre-load the texture data to the graphics card's local memory.

Features that set AGP apart from PCI

Probably the most important feature of AGP is DIME (direct memory execute). This gives AGP chips the capability to access main memory directly for the complex operations of texture mapping.
AGP provides the graphics card with two methods of directly accessing texture maps in system memory: pipelining and sideband addressing.
AGP makes multiple requests for data during a bus or memory access, while PCI makes one request, and does not make another until the data it requested has been transferred.
AGP doesn't share bandwidth with other devices, whereas the PCI bus does share bandwidth.

	AGP	PCI
1	Pipelined Requests	Non-pipelined
2	Address/Data de-multiplexed	Address/Data multiplexed
3	Peak at 533 MB/s in32 bits	Peak at 133 MB/s in 32 bits
4	Single target, single master	Multi target, multi master
5	Memory read/write only No other input/output operations	Link to entire system
6	High/low priority queues	No priority queues

What does all this mean?

DIME short for Direct Memory Execute, DIME allows for video card to use some of the main memory for texture memory with 3D graphics. Usually video cards have 4 MB of RAM, some have 8 MB of RAM, but DIME allows for 12, 16, or even more memory to be used by allocating some of the main system memory.

Pipelining As you should know from reading Hennessy and Patterson's 'great' Computer Architecture book, pipelining is an implementation technique whereby multiple instructions are overlapped in execution. A pipeline is just like an assembly line. There are various different steps (pipe stage or pipe segments) that contribute to the end result. Each of these steps are done in parallel. The opposite of a pipelined architecture is a sequential architecture, in which steps are completed sequentially or one after another, not in parallel.

Sideband
addressing the AGP bus uses sideband signals to send addressing information separately from data. This technique allows addresss informaton to be presented to the bus concurrent with a data transaction. The result is a more efficient use of the AGP bus for data transfers. With sideband addressing, AGP utilized 8 extra "sideband lines" which allow the graphics controller to issue new addresses and requests simulataneously while data continues to move from previous requests on the main 32 data/address wires.

Bandwidth the amount of data a network can transport in a certain period of time - it is the capacity for the rate of transfer, which is usually expressed in bits per second.

AGP's is pipelined therefore requests are executed in parallel, making execution faster than the non-pipelined PCI bus.
AGP's address/data is de-muxed, therefore the AGP pipeline can work with the data gotten from the de-mux. The PCI bus's address/data remained muxed so that the non-pipelined PCI bus works with the data gotten from the mux.

Further comparisons between AGP and PCI

AGP is a port (it only connects two nodes) while PCI is a bus
AGP does not replace the PCI bus, it is a dedicated connection that can be used only by the graphics subsystem
AGP and PCI also differ in terms of their minimum length and alignment requirements for transactions. AGP transaction are multiples of 8 bytes in length and are aligned on 8 byte boundaries, while PCI transactions must be multiples of 4 bytes and are aligned on 4 byte boundaries.

Benchmark Analysis
Benchmark - individual programs or a mixture of programs that are run on a target computer to measure the overall performance of the system, or to measure more specific aspects of the performance, such as graphics applications, I/O processing, etc. Any aspect of the computers performance that matters to the user can be benchmarked.

To find out if the AGP bus itself is responsible for better performance, it is necessary to use two identical cards. Actually, cards that are identical other than the different slot types.

Unreal Demo, by GT Interactive Software - used a time demo to measure the frames per second (FPS) of the 3D rendering:
Result: AGP = 18.36 FPS and PCI = 18.1 FPS

Windows 95 Benchmarks - 3D Winbench Large Texture Scene

AOPEN AX6L

FIC KL-6011

High resolution parts of the large textures have to be swapped to main memory, which has a large impact on the PCI system. We can see that AGP's DIME (direct memory execute) system handles it much better

Results

The current PCI bus supports a data transfer rate up to 133 MB/s, while AGP (at 66 mhz) supports up to 533 MB/s, which makes the AGP bus substantially faster.
The result of AGP is a much smoother frame rate and the ability to display 3D graphics and video that is many time more realistic and much higher quality than even before found on a PC.
AGP opens the door for high quality performance, high quality 3D by eliminating bottlenecks via direct access to system memory.

DIME	short for Direct Memory Execute, DIME allows for video card to use some of the main memory for texture memory with 3D graphics. Usually video cards have 4 MB of RAM, some have 8 MB of RAM, but DIME allows for 12, 16, or even more memory to be used by allocating some of the main system memory.
Pipelining	As you should know from reading Hennessy and Patterson's 'great' Computer Architecture book, pipelining is an implementation technique whereby multiple instructions are overlapped in execution. A pipeline is just like an assembly line. There are various different steps (pipe stage or pipe segments) that contribute to the end result. Each of these steps are done in parallel. The opposite of a pipelined architecture is a sequential architecture, in which steps are completed sequentially or one after another, not in parallel.
Sideband addressing	the AGP bus uses sideband signals to send addressing information separately from data. This technique allows addresss informaton to be presented to the bus concurrent with a data transaction. The result is a more efficient use of the AGP bus for data transfers. With sideband addressing, AGP utilized 8 extra "sideband lines" which allow the graphics controller to issue new addresses and requests simulataneously while data continues to move from previous requests on the main 32 data/address wires.
Bandwidth	the amount of data a network can transport in a certain period of time - it is the capacity for the rate of transfer, which is usually expressed in bits per second.