3DMark06 recycles some of the same models and scenes from the 05 version, making it more like an update than an all-new program. On the GPU front, 3DMark06 makes use of Shader Model 3.0, high-dynamic-range lighting, higher polygon counts, and a new shadowing algorithm. There are now four graphics demo scenes included; one of those is an all-new HDR-laden scene, while the others are 3DMark05 remixes.
The bigger changes may be in the CPU tests, where FutureMark has incorporated Ageia's PhysX API for physics and a multithreaded artificial intelligence routine. Both make good use of dual-core CPUs and multiprocessor configurations, and these CPU tests are now a component of the overall 3DMark06 score. FutureMark hopes its new overall score formula, which is determined by a mix of GPU and CPU component test results, will more closely map to overall system performance in future games than 3DMark05's largely graphics-based score.
That's the official line, at least. The day before 3DMark06's release proved to be an interesting one in Damage Labs. As I huddled here in seclusion testing all of the latest graphics cards on this new benchmark, a little green bird fluttered in and landed on my shoulder. It sang the praises of 3DMark06, saying that this new version should do a better job of tracking with real-world game performance than past editions, noting that 3DMark05 wasn't horrible, but was simply too vertex bound, as the example of the Radeon X1600 demonstrated. (The X1600 overachieved in 3DMark05, running relatively faster than in nearly any game.) 3DMark06, it cooed, should right that wrong and better achieve its stated goals.
Before it took off, though, the green birdie whispered that something unusual was going on with the PhysX API as implemented in 3DMark06. For whatever reason, said the unusually erudite avian creature, physics scales rather well on multi-core systems and the like, but makes no use of hardware physics accelerationvia a GPU or even Ageia's own PhysX cards, despite the fact that the program uses Ageia's PhysX API. My feathered friend was careful to use the phrase "hardware acceleration" for physics, being sure not to exclude GPUs from the possible sources of such acceleration. This was the first whiff I'd caught of a green-tinted take on physics acceleration, I was intrigued to note. It smelled of mint.
No sooner had the green bird flown out through the open window, no doubt seeking the shoulder of a pirate with eye patch and peg leg on which to land, than a redbird perched on the top of my monitor, singing a very different song. This creature, too, noted that 3DMark05 was excessively vertex-bound, but its take on the future was considerably grimmer. It warbled a tune of concern that 3DMark06 wouldn't track well with game performancea, erm, cardinal sin for "the gamer's benchmark."
The redbird then sang several verses of lament over specific issues with 3DMark06. The depth stencil format, explained the first sad stanza, requires a painful pixel-shader-based workaround on some hardware but not on others. Stanza two wistfully noted that no such shader-based workaround is required for the HDR graphics tests with AAthe tests are simply grayed out in the dialog box when antialiasing is enabled on hardware incapable of doing AA alongside 16-bit floating-point texturing and blending. (These were quite lengthy verses, in iambic pentameter.) The third stanza told the tragic tale of Shader Model 3.0 flow control, an incredibly important feature used only sparingly in 3DMark06, by the sound of it. Alas. Stanza four turned back to Shader Model 2.0, mournfully noting that the additional pixel shader load in these tests had a curiously green tint about itand yet did not improve image quality enough to justify the performance hit. The fifth verse was most melancholy of them all, acknowledging that things might have been different had the R520 not been late to market and unavailable during 3DMark06's incubation. Before I thought to quiz it on the PhysX question, the redbird fluttered away into the early evening, improvising a final stanza in praise of FutureMark's goals and expressing hope for 3DMark07.
Clearly, I need to cut back on the caffeine.