Since my interest was "upgrading" our Mac versus PC shootout, I tested the same graphics intensive apps on a G5/2.5 Power Mac with the two best graphics cards available: the GeForce 6800 Ultra DDL and Radeon X800 XT (8X AGP Dual-Link graphics cards).
SPECviewperf is a standardized graphics benchmark that's available in a Windows and Unix version. Since Mac OS X is Unix "on steroids," I compiled the Unix version using Mac developer tools (Xcode 1.5) with the help of instructions posted by Marcel Bresink. SPECviewperf simulates eight different real world apps. We chose the Maya animation simulation to graph.
The emphasis in SPECviewperf is polygons and lots of them. On the other hand, when you play a 3D Game, there is an emphasis on pixels and shading.
Halo is a great cross platform test since it uses advanced shading technologies. Notice we included results for dual nVidia graphics cards joined together in "SLI" mode.
LEGEND of GRAPHS
Xeon = Dual 3.4GHz Xeon (PCI-Express)
Athon = 64 Bit Athlon 3500+ (2.2GHz, PCI-Express)
G5 = Dual G5/2.5GHz Power Mac (8X AGP)
GeFU = nVidia GeForce 6800 Ultra (PC Edition - PCI-Express and Mac Edition - 8X AGP) GeFGT = nVidia GeForce 6800 GT (PC edition; PCI-Express)
QuFX = nVidia Quadro FX 3400 (PC Edition; PCI-Express)
FireGL = ATI FireGL V7100 (PC Edition; PCI-Express)
X850 = ATI X850 XT (PC Edition; PCI-Express)
X800 = ATI Radeon X800 XT (Mac Edition; 8X AGP)
SLI = nVidia's Scan Line Interleave (SLI) multi-gpu technology; two PCI-Express cards are linked together using jumper with SLI mode enabled in the driver.
The popular Doom 3 game will soon be released on the Mac, so we're adding it to our cross-platform suite of tests. We reran this test using the default "Demo1" instead of our captured sequence. Since none of the cards had more than 256MB of video memory, we ran the High Quality setting instead of Ultra.
We ran the "Inferno" map in the Unreal Tournament 2004 test, using the SantaDuck Toolpak for both Mac and PC. SLI mode speeds up the Flyby but does very little for the Botmatch as you can see below.
The Botmatch simulates game play more closely than the Flyby. Therefore the framerates are more in line with what you'll see when you are actually playing the game.
INSIGHTS and COMMENTS
1. Though the G5 Power Macs did well in the CPU tests, they were pretty much "smoked" in the 3D GRAPHICS tests. Part of it as to do with how much effort developers put into optimizing (or re-writing) the game code to take advantage of the unique features of the G5 Power Macs. Quake 3 Arena, though considered obsolete by avid gamers, is a perfect example of the potential of the G5 Power Macs to excel in gaming:
One of the lead programmers at Id Software (Graeme Devine) took a personal interest in optimizing the Mac version to take advantage of the Power Mac's Velocity Engine and Dual Processors.
2. The more expensive "workstation" (professional grade) graphics cards like the Quadro FX and FireGL beat all the "desktop" (consumer) graphics cards when running the polygon intensive SPECviewperf test. But the "desktop" GeForce 6800 Ultra was the "king" of 3D gaming.
3. Dual-Link DVI, as you know, is required for "9 megapixel" displays (up to 3840x2400 resolution) like the Apple 30" Cinema. The high-end Dual-Link Mac compatible graphics cards may seem expensive to Mac users at $499 for the Radeon X800 XT and $599 for the GeForce 6800 Ultra. But the PC Dual-Link cards we tested ranged in price from $850 to $1200. I was able to "Google" up a handful of GeForce 6800s with true Dual-Link DVI support. The lowest price was $670. All at once, the Mac boards seem like bargains.
4. We had two each of the Quadro FX 3400 and GeForce 6800 Ultra cards, so we were able to try out nVidia's SLI multi-gpu technology where you bridge two cards in adjoining PCI-Express slots. Once you enable the mode in the driver, your graphics processing sub-system now has dual processors. So do you get double the performance? No, but if the software is SLI "aware," you will get significant gains as you can see by the Doom 3 and Halo graphs above.
Those are respectable gains when you consider the bandwidth is shared and the memory of each card mirrors the other. In other words, there is no gain in bandwidth or memory capacity by linking the cards.
5. PCI-Express has a theoretical bus speed twice that of 8X AGP, but I postulate that the bandwidth advantage is under utilized. The current generation of graphics cards and motherboard designs don't even saturate a 4X AGP bus. We're hoping to set up a sub-test with two Windows PCs of identical specs and identical graphics cards except one will have PCI-Express and the other 8X AGP. That comparison should prove or refute our hypothesis.
