Choosing a Graphics Accelerator

Originally published in the February 1995 issue.

There are many ways to upgrade your computer. You can add a larger hard drive, more memory, or even a faster CPU if your motherboard allows. But the most dramatic thing you can do to your system is to upgrade your graphics card, because computing is a visual experience. It may not play the part in overall system performance that your PC's processor/memory and storage subsystems do, but the thrill of booting Windows in a high-resolution, true-color graphics mode for the first time is unmatched. By contrast, how much fun can you really have watching the boot sequence count up to 16MB after a memory upgrade?

The buzzword in the graphics industry right now is integration. Different components on graphics cards, such as graphics accelerators, clock generators, and RAMDACs (the digital-to-analog converters that send information to the screen), are being merged into single chips. Graphics cards--and, in some cases, graphics chips--are gaining the support for full-motion video formerly found on separate video-capture and -playback boards. At the same time, video-playback and 3-D graphics functions are being integrated into applications and operating systems software.

Combining parts reduces the cost and complexity of producing graphics cards, so prices are dropping rapidly. Putting additional functions on one card also frees up scarce motherboard slots and reduces the problems inherent in getting complex components to work together smoothly. In addition, providing dedicated horsepower on the graphics card means that the system's main CPU isn't taxed for demanding display tasks like full-motion video or 3-D animation.

That's good, because along with ever-higher graphics demands comes a quest for ever more graphics speed. Manufacturers are widening the path for data traveling between a card's accelerator chip and its onboard display memory from 32 bits to 64 bits and even to 128 bits. This, along with other architectural improvements, lets the graphics card handle more data per clock cycle, executing display instructions more quickly.

The width and speed of the input/output bus--be it ISA, EISA, or today's more popular VLB and PCI--also affect graphics performance. A card may be able to process data in 64-bit chunks, for instance, but it can only send it to your monitor in 32-bit chunks across the VLB or PCI bus. Still other speed factors are the card's display memory type, DRAM or the faster VRAM, and the efficiency of the specialized software drivers the manufacturer provides.

All these issues--accelerator design, bus type, the amount and type of video memory, and software support--play a part in choosing the card that'll set your screen ablaze. Read on to learn what to look for. (For more background on graphics-card design, see the sidebar, "Starting With Zero: How Graphics Accelerators Work.")

Basic Specifications

Technically, any graphics card you'll buy nowadays will be a fixed-function graphics accelerator, hardwired to take advantage of the application programming interfaces (APIs) built into Microsoft Windows, Windows NT, or IBM's OS/2. Accelerators have obliterated the market for slower, simpler frame-buffer cards, which leave screen-drawing duties to the system's main CPU. They've also replaced the first, more expensive generation of graphics speedups, boards with programmable graphics coprocessors. (The latter have been pushed into the back-office labs of graphics-chip vendors, where their programmability is useful in testing new algorithms and accelerator designs.)

The first step in choosing a graphics accelerator is to settle on a bus type. If you're looking to upgrade a PC that was built before 1992, you can pretty much forget about exotic 3-D and video options. You'll need to look for an ISA-bus card (unless you're unfortunate enough to have a Micro Channel-based IBM PS/2, orphaned by virtually every hardware vendor).

Most graphics vendors offer cards for 16-bit ISA slots, some of which perform surprisingly well for mainstream computing tasks like word processing or spreadsheet work. The market's main focus, however, has shifted to faster 32-bit local-bus solutions, divided between the competing VL-Bus and PCI standards.

Contrary to what you may have heard, the performance differences between PCI and VLB are relatively insignificant for graphics cards. Most modern 486-based PCs have VL-Bus slots, while most Pentium systems have PCI slots, though this is not a hard-and-fast rule. Some systems reverse that convention, while others offer both types of slots (although to date, their hybrid motherboards have been noted for relatively slow performance). In general, the industry is moving toward the PCI specification, and for that reason, the newest and hottest graphics cards appear first (and sometimes only) for PCI.

After catching the right bus, the next issue to resolve is the amount of display memory the card should have. The RAM required to form a frame buffer large enough to hold a complete screen's worth of data depends on the resolution and number of colors you desire.

At VGA resolution, for example, the display redraws 640x480 pixels each time the screen is refreshed. To support 256 simultaneous colors (8 bits per pixel) at this resolution, you need a frame buffer of 640x480x8 bits or 300K. A true-color (photographic quality or 16.7-million-color) display uses 24 bits per pixel, which equals 900K of data for each 640x480 screen. To see 24-bit color at 1,024x768 resolution, you need at least 2.25MB of display memory.

While a few cards provide the bare minimum memory needed for their targeted resolution (notably Media Vision's 2.25MB Pro Graphics 1024), vendors have settled on three standard memory levels for PC graphics cards: 1MB for 24-bit VGA, 2MB for 24-bit 800x600 resolution, and 4MB for true color at up to 1,280x1,024 resolution.

Realistically, 256 colors (8-bit color depth) are all you need for business computing tasks. If you're performing high-end multimedia, desktop publishing, or image editing tasks, however, a 16-bit (65,536-color) or preferably 24-bit palette, combined with high resolution such as 1,280x1,024, is essential.

While it's easiest to buy a card today with enough memory to support the graphics work you anticipate doing tomorrow, do-it-yourself upgrades for display memory are getting easier to install and cheaper to buy. Most manufacturers offer upgrades from 1MB to 2MB or from 2MB to 4MB, depending on the card, and upgrade prices--while still more than the difference between buying the board with one or the other amount of memory in the first place--are less exorbitant than they used to be. It's still likely, however, that you'll need to buy a proprietary memory module from the card maker rather than a generic one, as you can when adding SIMMs to a desktop PC.

VRAM or DRAM?

Some people refer to all graphics cards' memory with the abbreviation VRAM (video RAM), but that's not accurate. Genuine VRAM is a special type of memory optimized for use in video boards--it's dual-ported, which means that the chips can deliver their information and be reset for the next screen refresh in a single clock cycle. Standard DRAM (dynamic RAM) chips are single-ported, so the reset takes an additional clock cycle.

This means VRAM is inherently faster than DRAM, hence ideal for high-performance graphics cards. It's also much more expensive: You can expect to pay nearly $100 per megabyte for VRAM, compared to less than $40 for memory in DRAM-based boards.

Striving to keep costs down, chip makers and card vendors have gone to considerable lengths to boost DRAM's performance. One solution is page-mode interleaving, as used in Tseng Labs' W32i accelerator, which separates DRAM into two banks--one being used while the other is being reset. Another solution is what Chips & Technologies calls XRAM caching; that company's CT64300 uses a single fast-page-mode DRAM as a buffer to keep the main DRAM banks full of data.

Fast-page mode essentially cuts a couple of steps out of the data read/recharge cycle, while interleaving works on a higher level, accessing alternate memory banks without affecting the cycle itself. S3 offers fast-page-mode DRAM versions of most of its product line, including its latest 64-bit Vision864 accelerators. Tseng's W32p combines interleaving with fast-page-mode DRAMs.

For most uses, these souped-up DRAM cards perform nearly as well as more costly VRAM accelerators. DRAM graphics cards' relative performance begins to drop off in true-color modes, however, and they also can't keep up with their VRAM cousins' rapid refresh rates.

Most Refreshing

The longer you stare at a computer monitor, the more you'll appreciate the benefit of higher refresh rates or vertical scan frequencies. Support for more frequent screen redraws requires cooperation between your graphics card and CRT, which need to work in sync at a given combination of scan frequencies. Any VGA monitor can sustain a refresh rate of 60Hz, but prolonged viewing at that rate exposes a noticeable flicker that can tire your eyes. (We won't even consider card/monitor combinations that use interlacing, which cuts the effective refresh rate in half for guaranteed headaches.)

Newer multiscanning monitors--and any graphics card that you consider--should be able to handle a noninterlaced, 75Hz refresh rate at all resolutions up to 1,024x768 pixels, as recommended by the Video Electronics Standards Association. The limitations of DRAM force many DRAM-based cards to fall to a lower refresh rate such as 60Hz in true-color modes, so read the fine print in ads and specifications. Better-performing exceptions include most of the S3 Trio and 864-based cards, such as Diamond Multimedia's Stealth 64 DRAM, STB's PowerGraph 64, and Number Nine Computer Corp.'s #9GXE64, as well as ATI's Graphics Xpression.

By contrast, VRAM-based boards are nearly always able to sustain high refresh rates at all resolutions and color depths. Paired with compatible monitors, VRAM cards like Matrox's Impression Plus 220, ATI's Graphics Pro Turbo, and Number Nine's #9GXE64 Pro can provide refresh rates in the neighborhood of 100Hz at resolutions up to 1,280x1,024.

Besides reducing screen flicker below the eye's ability to discern it, ultra-high refresh rates provide a rich, saturated display. The result is bright and stable--once you've seen it, you'll never want to go back. As you'd expect, however, cards and monitors that support 100Hz or higher refresh rates extract a price penalty, so you need to balance your desire for a rock-steady display with a realistic assessment of the amount of time you'll be working with such stratospheric resolutions and color depths.

Stacking the Chips

So far, we've considered issues that influence the graphics card's output quality and speed, but performance begins with the graphics-accelerator chip itself. Although nearly every card on the market is based, as we said, on a fixed-function accelerator chip, these chips don't speed up every one of the possible Windows Graphics Device Interface (GDI) functions.

The first fixed-function accelerators worked only on the single most common GDI operation, bit block transfers (bitblts). Most current products address a range of Windows graphics issues, accelerating graphics primitives like pattern fills and line and polygon draws, as well as bitblts. Additional performance enhancements include support for a hardware cursor (in which the pointer shape is hardwired into the chip rather than created pixel by pixel), font caching, and linear addressing, which allows the system's CPU to access display memory directly for large image transfers.

Another factor to consider these days is whether the card integrates components such as the accelerator, clock generator, RAMDAC, and in some cases (or in the future) motion-video or 3-D support into a single chip. Cirrus Logic's CL-5434 and CL-5430 and S3's Trio764 and Trio732 chips are leading the way in providing integrated graphics solutions at the low end of the market.

By combining graphics acceleration, a clock generator, and a true-color RAMDAC in one chip, such silicon has lowered the price floor for respectable graphics performance. Cards based on these chips, like Diamond's SpeedStar 64 and Acronics' ASI 2000, typically offer 1MB of DRAM and 640x480-pixel true-color support for $100 to $150--perfect for entry-level systems and low-cost upgrades.

At the high end are new chips from S3 (Vision868 and Vision968) and ATI (Mach64) that combine graphics acceleration and full-motion video, and Matrox Graphics' MGA accelerators that support real-time 3-D Gouraud shading. Cards based on these chips include Diamond's Stealth 64 Video, ATI's Graphics Pro Turbo, and Matrox's Impression Plus.

Both full-screen video and 3-D graphics require functionality not originally built into Windows to be effective. Microsoft and Intel have jointly addressed video issues with their new Display Control Interface (DCI) specification, which adds a set of video-specific APIs to Windows' GDI functions (see Trends & Technology, November 1994, p. 55).

The 3-D picture is not quite so clear. Both OS/2 and Windows NT provide 3-D support through OpenGL, an industry-standard 3-D interface originally developed by Silicon Graphics, but no true standard has yet emerged under DOS or Windows. Competing APIs have been proffered by Intel (3DR) and Autodesk (HOOPS), and DOS-based game manufacturers have been lining up behind Criterion's RenderWare 3-D development library, Argonaut's ERender, and Reality Labs' Rendor Morphics, but the dust has yet to settle. (See the sidebar, "Entering the Third Dimension," for details.)

The Bandwidth Game: 32, 64, 128, Hike!

Bandwidth is defined as the amount of information that can move along a data pathway. There are two main pathways in your PC that affect graphics performance. The first, between the CPU and graphics card, is limited chiefly by the latter's bus connection. The 16-bit, 8MHz ISA bus limits the amount of data traffic to about 2.5MB/sec, while the 32-bit, CPU-clock-speed (typically 33MHz) VLB and PCI specifications allow roughly 30MB/sec of data to be transferred from the CPU.

The second data pathway runs between the card's accelerator chip and its display memory. Unlike the CPU-to-card route, which is constrained by the I/O bus interface, this road is wide open for graphics-card makers willing to stretch costs and memory speeds to their limit. Many vendors have doubled this onboard data path from 32 bits to 64 bits, with Number Nine the first to introduce a 128-bit graphics accelerator. The result is a bandwidth boost from roughly 100MB/sec in early 32-bit DRAM accelerators to nearly 500MB/sec for Number Nine's 128-bit VRAM-based Imagine 128.

Whether or not you need all this bandwidth depends on the work you intend to do. With a 256-color palette, a bandwidth of 200MB/sec is more than enough to handle the data received or requested by the CPU, screen refreshes, and graphics acceleration functions. In 24-bit mode at 1,024x768 or higher resolutions, you need nearly that amount just to handle screen refreshes and CPU traffic, leaving little room for acceleration functions. Under these conditions, lack of bandwidth causes a deterioration in graphics performance, as different subsystems are forced to compete for access to a limited data path.

The Software Side

The fanciest graphics card's advanced functionality is wasted when configured for Windows' generic VGA mode. Hence, driver software is perhaps the most important component of any graphics card. In systems, as on the highway, bad drivers are a major source of crashes, and less-than-optimal ones can have a dramatically negative effect on overall performance. The best drivers provide a stable computing environment, effective utilization of the graphics hardware, and--increasingly important as even low-cost cards grow adequately powerful--a slew of features to improve your productivity and promote ease of use.

Support for your favorite operating system (such as OS/2 or Windows NT in addition to regular Windows) is critical to your success with any given graphics card. So, in some cases, is support for specific applications--plenty of cards still come with outdated drivers for DOS text-mode programs like Lotus 1-2-3, but architects and engineers know to look for the latest, most feature-packed AutoCAD drivers.

Because Windows dominates today's computing scene, graphics vendors lavish the most attention on their Windows drivers. Many cards offer drivers that let you change graphics modes without resorting to Windows' quirky Setup program; scroll around a "virtual display" that's larger than your monitor's resolution; and use pan and zoom functions to navigate virtual screens or magnify fine details. Other features to look for include controls that let you adjust the size and color of the cursor; centering utilities that let you fine-tune your display for different graphics modes; and color-calibration facilities that help desktop publishers match onscreen colors with their output device. Some drivers even let you change resolutions and color depths on the fly, without shutting down and restarting Windows.

If your vendor's driver doesn't supply all these features, you might want to check out AnyView Professional, a $49 utility from Binar Graphics (1-800-228-0666). Besides letting you switch resolutions and palettes on the fly, AnyView Professional offers color-calibration, acceleration, and virtual-desktop functions, plus the ability to associate applications with particular display modes--switching Windows into 24-bit true-color mode when you launch your favorite image editor, then switching to a 256-color, maximum-resolution mode to see as many cells as possible in your spreadsheet.

Graphics drivers are subject to frequent revisions, updates, and occasional bug fixes, so make sure your vendor offers a strong support system, and check the company's BBS or online-service forum regularly to be sure you're using the latest version. Also, if you're really a stickler for upgrades, look for a card with a flash BIOS so you can upgrade it via software instead of having to pull out and plug in new hardware chips.

Video Connections

With multimedia applications climbing in popularity, you may want to consider including motion-video acceleration in your graphics-update plans. With the advent of the DCI extensions to Windows and video-accelerator chips from Video Logic, Weitek, S3, and Tseng Labs, full-screen, full-motion video is poised to break through the barriers that have so far given users the unappealing choice of postage-stamp-sized screens or costly, dedicated add-on cards.

Diamond Multimedia's Viper Pro Video card, for example, includes Weitek's P9130 hardware accelerator for full-screen, 30- frame-per-second digital video playback at resolutions up to 1,280x1,024, and retails for only $50 more than the otherwise identical Viper SE. Other video-aware products are becoming available in just about every price range.

But even if you opt not to indulge in a video accelerator now, you should make sure that the graphics card that you purchase can accommodate you if you decide to explore full-motion video in the future. The way to do that is to have the right connections--either a socket for a proprietary daughtercard or other add-on (which will limit your options but probably guarantee a conflict-free upgrade), or a VESA Advanced Feature Connector (VAFC) or VESA Media Channel (VM-Channel) connector.

While no single standard for connecting video components to graphics cards has yet emerged, the VAFC design provides a much higher bandwidth for multimedia data than the older VGA pass-throughconnector. The VM-Channel, a modular architecture robust enough to allow uncompressed video streams to be passed between the video and graphics controllers, is better still, though VM-Channel cards are generally more expensive than those with VAFC connectors. (For more information on the merger of graphics and video, see the sidebar, "Integration Now: Combining Graphics and Video Acceleration.")

Looking Good

Once a commodity market of generic EGA and VGA cards, the graphics arena has been sizzling ever since Windows converted computing from 80-column text to full-time, bitmapped graphics mode. With the market offering ever higher performance and ever more features at ever lower prices, there's never been a better time to upgrade your graphics system. Shop with your future as well as current needs in mind, and you'll be delighted at the difference from your same old screen.

[Back to Features and Top Stories|Computer Shopper Home Page]

Copyright 1995 Ziff Davis Publishing Company. All Rights Reserved. This material may not be reproduced in any form without permission.