By Jenny Donelan
Many battles have been lost and won in the graphics chip arena over the last half year, but none of them conclusively. And though the main combatants haven't changed-ATI, Nvidia, and 3Dlabs continue to vie for marketshare and benchmark ratings-the very measurements upon which these struggles are based have shifted. Since Computer Graphics World covered 3D chip vendors six months ago ("Graphic Equalizers," February 2002, pg. 26), the SPECviewperf benchmarks have been substantially updated to better reflect real-world performance. At the same time, the actual usefulness of benchmarks, always a controversial subject, is being questioned more than ever.
|The WildcatVP from 3Dlabs is the company's first card based on its new P10 graphics chip. |
Here's a synopsis of what has happened to date in the ratings: 3Dlabs' high-end Wildcat III 6110 card debuted with spectacular results this winter, but was subsequently trounced in most of the standard SPECviewperf tests by Nvidia's new Open GL-based Quadro4 900 XGL card. The Wildcat line, though pricey, had usually aced benchmarks in terms of sheer performance. 3Dlabs then began talking about the importance of "real-world benchmarks"-ones based on tasks that actual animators might perform with actual applications, instead of the "artificial" tasks created for SPECviewperf. Such talk sounded disingenuous, considering 3Dlabs' recent ratings. But other industry experts agreed, and shortly thereafter, viewperf 7.0 benchmarks appeared that did indeed more closely mimic real-world performance. Once again, 3Dlabs was at the top of the heap. But soon it shared that honor. As of May, both Nvidia and 3Dlabs showed best performance in different viewsets of both viewperf and apc benchmarks. (For information on the differences between the two sets, see "A brief description of benchmarks" below. ) ATI, for the most part, brought up the rear, though still with respectable performance at a good price.
But how important are benchmarks really? For example, 3Dlabs doesn't even bother submitting its highest end card, the Wildcat III 6210, for benchmark testing because the extra texture memory it features isn't measurable via benchmarks. In the words of Jon Peddie, president of Jon Peddie Research based in Tiburon, California, "No sooner did 3Dlabs introduce its killer Wildcat III with twin pipes and integrated pixel processor then ATI rolled out its Radeon 8800 and Nvidia trumped that with the GeForce 4 chipset, soon followed by its Quadro chip line. Does that mean if you bought a Wildcat III you'd be unhappy? Not really. Not if you need a huge memory pool for your textures, and not if you wanted real multi-sampling antialiasing the way it's meant in OpenGL. And if you bought an ATI Radeon 8800, would you be sorry you had after you saw the benchmark scores for the GeForce4? No, unless you run benchmarks for a living."
Although benchmarks remain a focal point for users and vendors alike, other issues are on the horizon. One of the biggest changes in the past six months has been the purchase of 3Dlabs by Creative Technology, a company known more for its consumer-level products. At the same time, 3Dlabs has rolled out the chip named the P10, which is neither OpenGL nor DirectX-based. It will work with DirectX 8 or 9 and OpenGL 2.0 when that ships, and will provide, according to the company, unprecedented power and flexibility. The first P10-based product 3Dlabs plans to ship is the WildcatVP. Other P10-based entry-level cards that Creative plans to issue this year are poised to take on "Longhorn," the forthcoming Windows 3D desktop. "This will be the killer 3D app, says 3Dlabs' Neil Trevett. "because [Longhorn] will be a big drain on resources."
Over at Nvidia, folks are pleased with the recent benchmark successes of the Quadro4 900 XGL and 700 XGL boards (For a review of the Quadro4 900 XGL, see pg. 60 of this issue.) Last year, Nvidia was a media darling on account of its programmable pixel and vertex shaders, introduced with the Direct X-based GeForce 3 chip. And it is now planning to build on that programmability with not only new products but also a new language-CG, co-developed with Microsoft-that will allow users further programming power at the pixel level. "It's a developer environment that's a compiler," says Nvidia's Jeff Brown, "And it will completely change the face of graphics."
Meanwhile, ATI is quietly readying its next generation of products. At press time, the company was preparing a major announcement. Despite the fact that the ATI boards have brought up the rear in recent testing, the company's director of workstations Ed Huang praises the SPECviewperf 7.0 updates. "It's a great new benchmark. It's more complex. Even though it's a synthetic benchmark, it's more representative."
With both 3Dlabs and Nvidia poised to offer "world-changing" new technologies, and ATI sounding confident and taking the high road, more performance battles are no doubt in the offing.
Jenny Donelan is Managing Editor for Computer Graphics World.
A brief description of benchmarks
The SPECviewperf benchmarks, which are written in C and designed to measure the 3D rendering performance of systems running under OpenGL, are updated and maintained by members of SPEC (mostly graphics hardware and software vendors). Currently, six viewsets make up viewperf: Each is designed to mirror operations within 3D applications such as 3ds max or Pro/Engineer, and each consists of 10 or so tests that mimic what the application might do in real life. Previous versions of viewperf, though based on applications, did not always reflect real-world performance, partly because they were based on applications that were no longer in general use, and partly because of the way the tests themselves worked. Viewperf 7.0 has been rewritten to test workstations and graphics cards in a manner more resembling that of actual OpenGL graphics applications. Still, because the SPECviewperf benchmarks are based on C, experts suggest, they lend themselves to the creation of vendor-made drivers that will enhance benchmark performance, but no other kind of performance.
To that end, another set of benchmarks exists as well, SPECapc, which are based on actual applications (such as 3ds max, Solid Edge, etc.) rather than written in C to mimic the operation of actual applications.
For more information on the benchmarks, and to view the latest results, visit www.spec.org/gpc. -JD