3D & 2D Graphics Controller Winners

By Donald Jenner

Half the price of your next Wintel workstation could be the cost of the graphics subsystem - the display adapter and the monitor. Prudent shopping can keep this cost element down; the heart of prudence is knowing what you need from the display system. Start with the display adapter.

A good display adapter, suited to CAD users' needs, costs anywhere from US$250 to US$2,500. The lower cost products are frequently lumped in with game-accelerators by commodity-box dealers without much graphics savvy; you need to walk in knowing which boards are likely to offer good price/performance value. High-end products involve such arcane specifications that buying them should only be undertaken in consultation with a specialist - whose interests had best be cognate with yours...

Two examples make this case; both have won accolades in recent months from sophisticated graphics users.

The last time CAD SYSTEMS looked at high end display systems, pride of place went to products based on 3Dlabs' Permedia and Glint chips. 3Dlabs' chips are still the foundation of some real winners, but Intergraph's Intense 3D Pro series is garnering more attention from the heavy-duty user crowd. These are people who need to move truly astounding amounts of graphics information to the screen in short order, to support realistic animation and similarly demanding 3D applications.

Intergraph has emerged as one of the top (if not the top) vendor of Wintel workstation offerings to this market, with its highly regarded TDZ systems. The Intense 3D Pro series of graphics accelerators (the 1000, listing for around US$1,600; the 2200, ranging up to US$2,400, depending on configuration) allow users buying systems from other vendors to add the graphics capabilities similar to those in TDZ workstations.

The low-cost Revolution 3D from Number Nine (US$170-US$300, depending on configuration) comes in versions for both PCI bus and Intel's AGP (Accelerated Graphics Port - see sidebar) architecture. It is based on Number Nine's 128bit T2R processor, a chip that is finding its way into a number of system-builder implementations as well as the companyÆs own boards. More data moves through the system - and a longer "word" instruction set means that multiple commands are passed to the memory bank in any given cycle.

List common features for both these processors, and they are remarkably similar:

Double-buffering means that one buffer stores what is on-screen, while the other is storing the next image as it is rendering. The buffers are switched back and forth. This results in smooth, flicker-free animation, significant in walk-throughs, and so on.

32bit Z-buffering: Design is increasingly 3D, even if display is still 2D. Z-buffering is used to manage hidden line removal, off-loading work from the central processor. The greater the bit-depth here, the finer the increments in depth rendering.

Perspective correction and fogging contribute to the realism of images. Things at a greater distance from the viewer are changed accordingly; texture elements converge and blur in a realistic way.

These and other 3D features are important, and both of these boards sport them. Equally important, both the high-end Intense 3D pro series and the Number Nine Revolution 3D models feature outstanding 2D performance. This has not been generally true for earlier designs. In particular, earlier display adapters based on Permedia chips were notoriously poorer in 2D performance than low-cost SVGA boards. Quite a lot of CAD is still 2D; the performance drop-off in those designs was a problem.

If there is that much similarity at the feature level, where are the advantages to the more expensive Intense 3D pro boards? When should you look more closely at the Revolution 3D?

In general, most CAD users will find a product like Number Nine's a better value. Even Intergraph makes this clear: The company promotes its TD225 systems (Intergraph calls these PCs, but they are priced and spec'd more or less the same as other vendorsÆ entry-level workstations) to the CAD/CAM market. The default board for these systems is the Matrox Millenium (based on Matrox's 64bit processor, priced similarly to the Revolution 3D, and with somewhat more modest performance characteristics as would be expected of an older design).

I concur in this view; in the newest Intergraph bench system here, I installed a Revolution 3D and saw about a 30 percent overall performance improvement. I found this board supported enough 3D capability for my needs, both in testing solutions and getting work done around the office.

CAD users buying systems from local system-builders, with the flexibility that entails, will certainly want to look at products like the Revolution 3D. Integrating these boards is more likely to be something a local system-builder has done before. There will be fewer problems.

Older systems, getting a bit ho-hum performance from the latest software revs, but not quite ready for replacement, could also benefit from a graphics-board upgrade. This is a tricky matter; you don't want to put a lot of money into an older system. That Number Nine has a sub-US$200 version of the Revolution 3D makes thinking short-term upgrade (and a couple years' extra life on the systemà) viable. [Keep the old board; when it comes time to upgrade, you may want to swap the old board back in and give the old box to a local favorite charity, while adding memory to the Number Nine and having it installed in the new system. It is that good a design, I think.]

Who needs the high-end Intense 3D pro? Where engineers, in particular, are running advanced modeling and simulation systems, throwing hardware at the speed problem makes sense.

Users in this class probably should look at a well-integrated solution where the display hardware is well-tuned to the rest of the hardware as part of the design. This has always been part of the "workstationö concept, and it is clearly what Intergraph does in its TDZ workstations. If that is not an option, for whatever reason, a product like the Intense 3D Pro offers that class of performance, well tuned to OpenGL software.

The central processor in most computers communicates with peripherals, including the graphics display system, over a bus. Make the processor as fast as you want, data can only move as fast as the bus allows. Put a blazing processor on a system with a slow bus connecting to the graphics display system and graphics will be substantially less impressive.

Workstation vendors, who have very little to sell these days but performance, have addressed this in two ways.

In its O2 low-cost workstations, Silicon Graphics adopted a strategy where some of its peripheral systems shared direct access to the main system memory with the central processor. Graphics data never moved across the bus; it was just shared. This improved performance in these carefully tuned, well-integrated systems.

At a higher level - in SGI's Octane workstations and some of Sun's UltraSPARC boxes, e. g. - the bus was retained for less speed-sensitive peripherals. Switching was used to manage things like graphics subsystems. This let the processor have, effectively, a dedicated connection on demand with the speed-sensitive peripheral.

Intel adapted the lower cost, lower end solution of shared memory for some of its newer motherboard designs. Some companys, Number Nine among them, have adopted the AGP specification, and offer tuned versions of their boards to take advantage of it.

The merits of using this shared-memory connection are subject to dispute. A couple of the issues are particularly significant for CAD users.

First, this is a Win95/98 solution only. For Windows 95, things are rough; Windows 98 - when Microsoft finally finishes it (current guess: late-June '98)... - will have built-in support. Thus far, WinNT support is simply not available for most AGP graphics adapters (Number Nine appears to be an exception; note well, though, that WinNT has a reputation for being unfriendly to driver additions).

Second, while there is a performance gain, it is not a great gain. In some cases, the gain is as much a factor of faster display system memory, as it is of the improved connection to the processor.

Intel has taken an interesting direction with AGP, but this is not the company's forte. When the operating system support is in place, and after a generation or two has passed, AGP will become more interesting, perhaps.