Tech-Trek[AGP]

Home ] Up ] Future Fantastica ] [ Tech-Trek[AGP] ] Hard-a-ware[video memories] ] Hard-a-ware[LCDs] ] Inventions[Maglevs] ] Mathematica[Humour] ] Surfs Up ] Crossword ] Soft-a-ware ] Trivia ]

emailwhite1.gif (13819 bytes)

search.gif (12172 bytes)

addchannel.gif (1055 bytes)

Submit an Article

AGP -- what's in it for you?

The graphics intensiveness of today's business applications is increasing. In place of mere words, we have become more reliant on images, icons, and charts. Add entertainment and educational software to the mix, and you have a graphics environment fit for the new millenium. Entertainment and educational software heavily stresses the graphics environment, of which software developers in this industry are acutely aware. The folks at Intel know this, too. They know that as graphics content increases, so does the need to process that content. With the advent of realistic gaming scenes, complete with clouds and fog, the entire industry has realized that the current graphics interface, the Peripheral Connection Interface (PCI), is insufficient. In response, Intel designed the highly publicized--and now controversial--graphics interface, the Accelerated Graphics Port (AGP). AGP uses main PC memory to hold those large 3-D scenes, effectively giving systems unlimited graphics memory. That could yield more attractive graphics and hopefully higher performance. As part of AGP, Intel has also created a dedicated port and a path from the graphics board to the PC's main memory. So we have a nice, tidy package. In theory, AGP certainly promises groundbreaking performance boosts, but in the real world, AGP seems to offer only limited improvement in the gaming area. So the question now is: Does AGP really improve graphics performance overall? And does it deliver, in a universal view, what it promised?

 

On the Bus : Several years ago, the industry recognized that graphics would become the bottleneck for most desktop systems, and so vendors began working on various solutions. First came the Video Electronics Standards Association (VESA) VL bus, an extension of the existing ISA bus. VESA, which represents graphics, card developers and the PC industry's big three (IBM Corp., Intel Corp., and Microsoft Corp.), worked to develop the new standard. Though the VL bus did what it was intended to do (mainly, ease the burden on the ISA bus), it couldn't handle increasingly complex graphics information. Peripheral vendors also saw great promise in the VL bus because it was speedier than the ISA bus, and they began to develop VL-based SCSI cards and other disk controllers. This only served to increase the congestion on the ISA bus and led to no effective performance improvements. The VL bus lasted only about a year before Intel and the graphics industry replaced it with PCI. PCI provides a separate bus that isolates the PCI data stream and relieves ISA bus congestion. PCI also delivers double the bandwidth of the ISA bus, and the 66MHz of bandwidth it provides helps move data from the CPU to the graphics card quickly. As with the VL bus, peripheral vendors saw a separate, high-speed bus that they could use for disk controllers and network cards. But unlike its predecessor, PCI has not suffered from the addition of these non-graphics peripheral devices. Graphics information still moves quickly enough to satisfy our 2-D needs. Yet PCI faces the same demon that dogged the VL bus: increasingly complex graphics information, particularly 3-D graphics. PCI works marvelously with standard 2-D (x- and y-buffer information) and general business graphics, such as those used in Windows 95 or NT and most games. But it is severely challenged by intense 3-D graphics. To render 3-D graphics, the graphics controller must manage texture data and z-buffer information. Texture data adds a more realistic look to a scene, like the folds in a cloud, for example; z-buffer information provides depth. Each of these additions takes gobs of memory, and unfortunately both use the same memory space. So graphics board vendors face the trade-off of better texture vs. increased depth.

 

Tripped Up in Real Time To end this dilemma, Intel rolled out AGP in July of 1996. AGP has the potential to quadruple the bandwidth of PCI to 528MB/sec and, most important, uses main memory for textures and the z-buffer. This provides a fairly fast pathway, while eliminating the texture vs. z-buffer memory contention of PCI. The speedier AGP bus also opens the door for real-time 3-D graphics, something that would likely choke the PCI bus. And even though Intel uses standard PCI calls for AGP, it streamlines PCI memory storage to eliminate performance lags. Lifelike or realistic 3-D scene creation requires intensive three-dimensional geometric calculations, generally handled by the CPU. The graphics controller processes the texture data and texture bitmaps, called texture maps; this is where the difficulty begins. The graphics controller must fetch elements from as many as eight texture maps and average them out to produce one pixel of the final scene. Once the pixel is produced, the graphics controller must write it to the buffer. But because texture maps are so large, they cannot reside in the graphics board's frame buffer. Instead, they reside in the main system memory. The philosophy behind this arrangement is that main memory is cheaper than dedicated graphics memory; therefore, users will likely have sufficient amounts of main memory available for use. To access the texture maps in main memory more quickly, AGP uses a technique that Intel calls Direct Memory Execute (DIME). DIME effectively connects main memory to the AGP/PCI chip set, allowing speedy access to texture data. DIME lets the graphics controller access texture data in main memory, where data capacity is limited only by the size of the main memory itself. Without DIME, the graphics controller can access and manipulate only the texture data that fits in graphics memory. That limited amount of data is insufficient for larger textures. Managing this entire process is the Graphics Address Remapping Table (GART); GART is built into PCI/AGP chip sets. Standard AGP transmits data on the rise and fall of a clock cycle, meaning that with each clock transition, AGP transmits data. For this reason, Intel calls it 2x AGP. Since one clock cycle has four transitions, standard 2x AGP transmits four times the data of PCI. As the clock cycles from a low voltage (a digital 0) to a high voltage (digital 1) the AGP chip set transmits data. Once a high voltage is achieved, the chip set transmits again. The chip set will transmit again with the fall of the clock from a digital 1 to a digital 0, and again when a low voltage is sensed for a total of four transmissions per complete clock cycle. With timing modulation, Intel expects to double AGP's current throughput to more than 1GB per second and will call it 4x AGP. Like PCI, AGP uses a 32-bit connector. But AGP's connector, which resembles the now dead Micro Channel Adapter, has 64 contacts and a 64-bit wide bus. These additional contacts enable pipelining and queuing of requests, effectively eliminating latency. In addition, AGP has eight extra sideband address lines that allow the graphics controller to issue simultaneous commands--called sideband addressing, or SBA--while processing the main 32 data/address wires. So though both AGP and PCI can burst data, the pipelining capability of AGP will allow it to achieve higher sustained throughput rates than PCI.

Work it out with the Motherboard : To support AGP fully, a motherboard must use Intel's new 440LX PCI/AGP chip set, currently found only on Pentium II-based motherboards. The new chip set supports AGP and Intel's Quad Port Acceleration (QPA), which increases system bandwidth. SDRAM support yields faster memory access, an imperative for AGP. The new chip set also offers advanced power management and Ultra DMA/33 for higher-speed IDE devices. The non-AGP features alone should improve system performance, and when combined with AGP, should produce a very speedy system indeed. By early 1998 we expect to see the 440LX PCI/AGP chip set on motherboards using non-Intel based processors. That should broaden AGP's acceptance. Of course, AGP requires some form of software support from the OS and from graphics drivers. According to Microsoft you'll need Windows 95 OEM Service Relase 2.1 or a patch program, USBSUPP.EXE, that includes the Universal Serial Bus (USB) supplements. Your current Windows 95 PCI graphics-device driver will support an AGP graphics card, but you'll also need DirectX5, which is the only version to support DIME. You'll also need a virtual device driver called VGARTD.VXD, which should be part of your graphics driver installation. This VXD turns on the DIME feature and, as the name implies, is a virtual GART driver. With every giant step forward, something is left behind; in this case it is the current crop of PCs that fall by the wayside. AGP is a completely new bus with a physically different connector and chip set. And due to the nature of AGP, it requires an extremely powerful processor to make the whole thing work. This means that there is no way to upgrade an existing PC to AGP, short of replacing the motherboard, the processor, and of course the graphics card. But is AGP worth the extra effort? Intel surely thinks so, as do graphics card vendors who stand to gain from the widespread acceptance of AGP. Several graphics card vendors, including ATI Technologies, Diamond Multimedia Systems, Matrox Graphics, NVidia, Number Nine Visual Technology Group, and STB Systems, have introduced AGP versions of their PCI-based graphics cards. But even though these graphics cards all support AGP, they aren't created equal. Aside from graphics controller capabilities, each vendor is free to enable either the complete set of 3-D effects or a subset thereof, with a few choosing to implement some effects in software. By using a software approach, vendors avoid having to reengineer their graphics controllers. Even better for them, they can rush a product to market. But software implementation of 3-D functions degrades graphics performance. Each of these graphics cards includes a substantial amount of onboard RAM, with both Diamond and Number Nine using as much as 8MB standard. Nearly all of the graphics cards can support a maximum of 16MB of RAM, which makes for a very large frame buffer and enables the storage of frequently used texture maps. An 8MB frame buffer is large enough to cache 64 textures at 256x256 pixels and a 16-bit color depth. With 16MB of local RAM, you can easily cache these textures, plus double frame buffers and the z-buffer for depth data, and still have cache space left over. Clearly, a large frame buffer will benefit 3-D performance, and as we all know, it will also benefit standard 2-D performance.

The Real Test Tests reveal no real performance improvements for AGP with running standard business applications. Business Winstone 97 scores, for example, reflected no real difference between PCI and AGP boards, indicating that AGP has no effect on standard 2-D performance. Similarly, there was no significant improvement in 3-D performance as measured by 3D WinBench 97. We saw a big performance improvement when we used the 3-D Large Texture test, but only with some of the boards we tested. Graphics boards such as the ATI Xpert@Play, which supports AGP texturing, the full use of main memory, GART, and DIME, exhibited a sevenfold performance jump. This indicates that AGP does in fact provide a performance edge with large textures. The Number Nine Revolution 3D board, on the other hand, which supports only local texturing--meaning it uses the frame buffer for all texture information--had a nearly flat performance curve compared with the PCI version of the same board. Currently, only games use 3-D textures, but Intel believes that more applications will soon follow once developers become comfortable with 3-D and realize its potential. Microsoft plans to include support for AGP, including GART and DIME, in Windows 98 and Windows NT 5.0, which should add to AGP's universal appeal. Only time and widespread adoption will really indicate AGP's true worth. Our impression is that it is definitely worthwhile. But as with MMX technology, which has been around for almost two years, the market must recognize the need for it. That means software developers must find a way to make 3-D graphics work for business applications if AGP is to move beyond its uses for 3-D gaming. Otherwise, AGP will meet the need of only a relatively small group of PC users.

In Bombay, there are a whole lot of AGP cards available from Asus,SiS,Cirrus Logic,Trident,Prolink,Matrox etc. The best buy though costly at around 10 grand is the Matrox G200. Normal AGP cards like the one based on the Intel i740 chipset are available for Rs.2,300 to Rs.3,500 for 4MB and 8MB VRAM respectively. — Saumitra D9

Back Next

Copyright©Samsoft Technologies®1998 for SFE WebAuthor:Saumitra.M.Das [D9] Hits: Hit Counter

suggest a site   sign our guestbook