Compaq AP750 - Professional - 256 MB RAM Manual page 10

Graphics: The Workstation Difference
The implementation of AGP as a point to point bus enables much of AGP's performance. By
eliminating multiple devices on the bus, signal integrity problems are greatly reduced and bus
clock rates can be increased. Likewise, eliminating the need to determine which of several
devices has control of the bus simplifies the software interface and improves performance there.
Another advantage of point to point connection is that the graphics device has the full bandwidth
of the AGP bus -- it doesn't share this bandwidth with any other devices.
AGP is implemented to provide a high bandwidth, low overhead connection between the graphics
device and system memory. This connection, combined with DMA transfers, allows effective
and efficient communication between the system and graphics.
The major disadvantage of the point to point connection of AGP is that it only allows a single
graphics device. You can't implement multihead displays with AGP based graphics. System
vendors have overcome this problem by using both AGP and PCI graphics devices in a single
system.
The major impetus for the creation of AGP was to allow textures for 3D graphics to be stored in
system memory. As stated in the AGP specification:
"In general, 3D rendering has a voracious appetite for memory bandwidth, and continues
to put upward pressure on memory footprint as well. As 3D hardware and software
become more pervasive, these two trends are likely to accelerate, requiring high speed
access to ever larger amounts of memory, thus raising the bill of material costs for 3D
enables platforms. Containing these costs while enabling performance improvements is
the primary motivation for the A.G.P. By providing up to an order of magnitude
bandwidth improvement between the graphics accelerator and system memory, some of
the 3D rendering data structures may be effectively shifted into main memory, relieving
the pressure to increase the cost of the local graphics memory. Texture data are the first
structures targeted for shifting to system memory... "
Using main memory for textures is referred to as DIME texturing -- Direct In Memory Execution
of Textures. DIME texturing supports large textures, but has two significant costs compared to
dedicated texture memory on the graphics accelerator: lower graphics performance and system
performance impacts caused by sharing main memory between graphics operations and CPU
operations.
At the time AGP was created, memory prices were in the $40-$50 per MB range. Memory prices
today are in the $1-$2 range. Thus, some of the original economic justification for AGP is
greatly reduced. Low cost graphics cards with 16 MB to 32 MB of memory are becoming
common.
0054-0499-A
10