Hi Paul and ALL, Article...Intel i740 processor promises to rewrite the marketplace..
A train wreck ahead in 3-D graphics chips?
February 10, 1998
Electronic Engineering Times via Individual Inc. : Santa Clara, Calif. - With the introduction of Intel Corp.'s i740 graphics processor rumored to be only weeks away, the stage is set for a collision of epic proportions in 3-D graphics. Intel's powerful chip will slam head-first into a pack of " Free-D" chip vendors scrambling to move upscale, just as game- oriented chip vendors-both established powers and dynamic startups-are trying to force their way into the mainstream of PC graphics. And in an irony suitable for the always ironic PC industry, the whole brawl may be made moot by the collapse of the market for midrange PCs.
In the midst of such turmoil, Intel's introduction threatens nothing less than the end of the graphics market as we know it, according to industry executives. "I think that when the dust settles, there will be a high-end market for games. There will be a low-end market that uses sub-$10 graphics chips. And there will be Intel," said Andy Keane, vice president of marketing at high-end vendor 3Dfx Interactive Inc. (San Jose, Calif.)
But before the market reaches any kind of stability, it is destined first to splinter, breaking along a number of well-known fault lines.
The first of these is applications. Today, only high-end games and some CAD tools use the full performance of top-of-the-line 3-D chip sets. These segments are small enough to ensure that the high-end chips will sell mostly in the premium-priced add-in board market. Motherboard vendors can't absorb the difference between $35 for a high-end chip and $10 for a so-called Free-D chip, one that folds a simple 3-D rendering engine into what is essentially a VGA graphics accelerator.
Once out of these specialized markets, though, there virtually are no applications for moderate-performance 3-D. "We make a joke that there are only two 3-D applications: One is games, and the other is 3-D Winbench," said Jon Peddie, president of 3-D market watcher Jon Peddie Associates (Tiburon, Calif.). Business applications, such as presentations and real-estate walk- throughs, should start appearing later this year, he said, but for now the market is negligible.
And that is a problem all 3-D vendors outside the high end must face. With no compelling applications, PC buyers haven't been willing to pay extra for 3-D hardware. And with a cost ceiling of zero, Free-D chip vendors have been unable to provide the kind of performance that would attract even existing games applications. Game developers would rather rely on software rendering than work to exploit the small performance improvements of Free-D.
"While many hardcore gamers have great 3-D accelerators like 3Dfx's Voodoo chip set, the large majority of consumers do not have capable 3-D hardware," said Tim Sweeney, founder and chairman of Epic MegaGames Inc. (Rockville, Md.). "So, we are optimizing [our games] for two separate targets: hardware rendering on PCs with 3Dfx-class accelerators such as Voodoo and [Nvidia's] Riva 128; and software rendering on PCs with only MMX."
That doesn't leave much market in the middle.
Political correctness
The second fault line in 3-D concerns the applications interface. All the successful high-end 3-D chips use a single approach to drawing 3-D objects, based on the graphics pipeline pioneered by Silicon Graphics Inc. (Mountain View, Calif.). Surfaces are broken up into triangles-a process called tessellation-and then each individual triangle is projected onto a viewing plane. Each triangle is lit, textures and possible bumps are mapped onto it, and features like fog are introduced. Then the pixels that make up each triangle are written into the frame buffer.
There are many other ways of drawing a realistic scene-some of them radical, such as the NEC VR's use of infinite planes or Nvidia's early use of quadratic patches. Other techniques, such as ray tracing, are well known but beyond the reach of PC graphics right now.
But the SGI-style approach has donned the cloak of orthodoxy in the games world. A new approach, even one that provides significantly better performance, will meet with hostility from developers who have learned to make music with the existing tools.
"It is very hard to change the industry infrastructure," Keane said. "Programmers understand a certain way of doing things, and they don't want to change. Look at SGI-they have the budget to do anything they want, but they still draw triangles."
This orthodoxy tends to split the 3-D world into two neat segments: chips- including the i740-that do things the orthodox way, and chips that don't. The latter either adapt or disappear.
One item that is not in the orthodox canon, by the way, is the actual application programming interface. Early on, Microsoft Corp. attempted to drive all PC 3-D chips to adopt its Direct3D API. This has pretty well worked in the Free-D segment, where there aren't any applications anyway. But in games, where the quality of the gaming experience depends on wringing out every drop of chip performance and feature, developers still cling to vendor-specific APIs. In fact, the most sophisticated game makers go beyond the vendor's API and become intimate with the entrails of the chips. "We have some developers who know more about the Voodoo chips than anyone here except our chief architect," Keane said.
Geometry matters
That leaves one more axis for differentiation among chips: how they go about performing the computations in the orthodox pipeline. Here major differences will appear in 1998.
The biggest issues occur right at the beginning of the pipeline. Most chips and chip sets today let the Pentium CPU handle not only the game mechanics, but the process of breaking surfaces into triangles, projecting the triangles onto a plane and sorting them to determine which are visible.
The tessellation process usually only happens once per solid object, so that's not a big deal. But projecting the triangles onto a plane-the so-called perspective transform-requires multiplying the coordinates of each vertex by a 4 x 4 matrix. These operations are generally done in floating-point format to preserve accuracy.
As rendering has grown faster, transforms on the CPU have become a bottleneck. This has led most vendors to design-at least in the laboratory-dedicated floating-point transform chips or cores that could take on the task.
But before it could become a trend, the move ran afoul of one of Intel's sacred business principles: Keep everything you can on the CPU. Intel improved the speed of its processors for single-precision floating point, and now is promising to multiply that speed considerably by permitting parallel, single- precision multiplies on the MMX-2 hardware of the Katmai processors, a spin of the Pentium II slated for early 1999.
At the same time, Intel is moving to make sure there is a strong market presence for 3-D chips that rely on Katmai. Enter the i740.
With performance near or slightly above existing high-end chips and with an attractive price point believed to be in the $35 range, the new device is calculated to make life difficult for any mainstream chip that spends silicon on geometry processing.
Few think chip makers can add floating-point units to consumer chips running on Windows 95 PCs and still come in below $35. In any case, Microsoft is closing the midrange market to this option for at least a year, since it has decided not to support dedicated hardware for geometry processing in its pending 6.0 release of the Direct3D API, slated to ship by June.
"When we looked at the consumer segment, we didn't see anyone with plans for dedicated geometry-processing hardware in the summer '98 [to] '99 time frame," said Kevin Bachus, product manager for DirectX at Microsoft (Redmond, Wash.). "Software developers tend to use their own geometry-processing code. We are rewriting the geometry code in DirectX now in an effort to get them to use our code running on the CPU. We are writing it in such a way that we can easily redirect any calls to our code to dedicated hardware, which we expect to support in DirectX 7.0." That version should ship in the summer of 1999, Bachus said.
Even backers of dedicated geometry processors are circumspect. "There's no doubt hardware support for geometry processing is critical in the workstation space," said Neil Trevett, vice president of marketing for 3Dlabs Inc. (San Jose, Calif.), which formally rolled out its geometry chip, called Gamma, last week. However, Trevett said it's not clear whether dedicated geometry processing would ever be affordable for consumer-class PCs.
If geometry processing has to stay on the CPU, that leaves only a couple of areas for differentiation. One of these, clearly, is the rendering pipeline. But rendering algorithms are sufficiently understood that no one has been able to make much of a breakthrough here. Designers only get their chips to render faster by adding more pipelines. And that may not be much differentiation.
Intel hoped the i740 would offer better than twice the performance of its nearest competitors by dint of its parallel 3-D graphics. As it turned out, though, the i740 is, at best, only a tad better than the leading chips from companies like 3Dfx, Nvidia and Rendition.
"The i740 is not the slam-dunk Intel was planning on," said Dean McCarron, principal of Mercury Research (Scottsdale, Ariz.). "Compared to just first- and second-generation 3-D chips from S3 and others they are close to their goal, but others also came out with a parallel-pipeline implementation that blew them [the second-generation chips] out of the water. "
That leaves one big area for differentiation that has been very little explored: the process of sorting the triangles to determine which ones you actually need to display in a scene and which ones will be obscured.
This process has long been the secret sauce of game developers. It was handled in software, and the potential benefits were huge: It might be possible to ignore, say, three- quarters of the triangles in a scene most of the time. Developers became masters of triangle ordering.
Hardware shift
But in the next generation of hardware, that task may shift to hardware. Two new ventures, GigaPixel Corp. (Santa Clara) and Raycer Graphics (Palo Alto, Calif.), are independently working to put triangle-ordering algorithms in hardware.
GigaPixel, started by Compcore Multimedia founder George Haber, has developed a proprietary sorting algorithm that it intends to include in a licensable 3-D graphics core. Haber said that the chip itself selects only the visible pixels to render, allowing application developers to focus their efforts on game mechanics and artistry. Like many new designs, the GigaPixel approach uses chunking, a partitioning technique to reduce off-chip bandwidth requirements. Borrowed from Microsoft's ill-fated Talisman architecture, the technique-also called tiling-is said to improve throughput and the appearance of images.
Raycer, which would not comment on product development, is reportedly planning a 3-D chip set, rather than a core. The company is employing a sorting algorithm originally developed for the Search for Extraterrestrial Intelligence (SETI) program. SETI, like a 3-D game, needed to sort through huge volumes of mostly irrelevant data. Raycer is said to have found enough commonality between the two problems to dramatically improve 3-D performance.
But both companies will have to overcome remaining barriers. One may be the same hurdle that defeated quadratic patches-the unwillingness of game developers to give up their secret sauce, even for more performance.
The other problem is not limited to the new ventures, but will be faced by all the players in the impending collision-the reality of today's PC market. The only major growth in recent months has come from under-$1,000 systems- emphatically not the kind with big displays, sizzling CPUs and high-end graphics cards. These penny-pinchers are certainly not a promising market for a $35 3-D chip.
With no compelling applications other than high-end games, all of the players rushing into the new 3-D performance space may face the same ugly uncertainty: There may be no large market for their wares. If that proves true, the impending train wreck could leave the high end relatively unscathed, the under- $10 market intact and Intel dominating a dwindling space in the middle. It could be a hollow triumph for the i740.
______________________________________________________________________
Regards, Michael |
