3D Mark 2006 shows us a number of things right off the bat: to start with, we see how old the 6600 GT really is. The once-legendary midrange champion is looking completely outclassed in this newer benchmark, and it looks like it might be time to retire the old girl. The 2600 XT and X1950 PRO seem to be performing well, but take a close look at the AGP 3850 and the underclocked PCI Express 3870: the scores are very, very close. Once again, the difference between AGP 8x and PCIe 16x doesn’t seem to be a heck of a lot.
But synthetic benchmarks are only so useful. Let’s see how some actual games behave, starting with the Doom3-engine based Prey.
Before we dig into this benchmark, it’s important to note why were using this older game to represent OpenGL, instead of the newer “Quake Wars”; the Radeon 3850 AGP couldn’t run the latter without crashing. This was odd because all of the other Radeons, including the X1950 PRO AGP and 3870 PCIe, could run Quake Wars no problem.
We have heard rumors that the Omega drivers might allow the 3850 to run Quake Wars, but we weren’t about to go switching drivers for each individual game, so we went with Prey. Interestingly enough, the X1950 PRO AGP couldn’t run Prey without crashing; clearly there are some bugs in the old AGP Radeon drivers that need some attention.
Prey shows us a solid ramp-up with each video card tier. It’s interesting to see the old Geforce 6600 GT perform so admirably with 4xAA enabled: clearly it could have handled this game easily with no AA.
While both of the low-end cards perform well at 1024x768, ramping up resolution quickly slows down the 6600 GT and 2600 XT — both of these cards are unplayable above 1280x1024. Contrast this with the AGP 3850 and PCIe 3870, which stay at around 30 frames per second regardless of resolution; what we’re seeing here is a severe CPU bottleneck, but no real evidence of an AGP bottleneck compared to PCIe.
It will be very interesting to see if the dual core CPU we get for part 2 of the AGP analysis will show the same bottleneck. Let’s move on to the toughest game of them all: Crysis.
For this first Crysis benchmark, we set all details to “Medium”, except for Physics which was set to “Low”. When we tried setting Physics to “Medium”, the single core Athlon 64 took a nosedive and could only muster an average 13 frames per second at 1024x768 with the Radeon 3850. This shows that physics requires a lot of computation power — probably multithreaded — and we will be very interested to see if this trend continues in part 2 of the AGP analysis when we have a dual-core CPU to play with.
With Physics set to Low and the resolution at 1024x768, Crysis is playable on all of the test cards except for the 6600 GT. The 2600 XT, X1950 PRO, and both Radeon 3800 series cards can pull playable — if not smooth — frame rates at this resolution. The PCIe Radeon 3870 shows a small but notable lead over the 3850, but that is probably attributable to the higher memory speed as opposed to the PCIe bus. As resolution increases the X1950 PRO hangs in there but slows down a bit, and the 3800 cards are in a dead heat.
Clearly we’re seeing more of the same CPU bottlenecking. This game requires some real CPU muscle to perform, and the powerful Radeon 3800 cards are slowed to almost the same speed as the older X1950 PRO. Keep in mind that the X1950 PRO doesn’t have DirectX 10 capabilities, and it isn’t being put through quite the same workout as the 3800 series cards, although it still produces excellent visuals; it’s difficult to see any difference.
Let’s see what happens when we up the ante and increase the shader and texture details to the High settings in Crysis: this should shift the bottleneck from the CPU somewhat, back to the video cards.
In this benchmark, Shader and texture settings are set to “High”, Physics is still set to “Low” and the remaining settings remain at “Medium”. With the increased demand on the graphics card, the CPU bottleneck becomes less prevalent, and the lesser cards are really slowed down once we hit 1280x1024 resolution and above. The underclocked 3870 PCIe shows a solid lead over the 3850 AGP, but once again this is probably attributable to the 3870’s higher memory speed and not the AGP bus. When we reach the 1920x1200 resolution, both 3800 cards are equal again, although none of the cards are really that playable at these settings (average frame rates hover below 20 frames per second).
Let’s move on to a game that’s very visually appealing, but much less demanding on the hardware when compared to Crysis: Unreal Tournament 3.