The Performance Of Voodoo2 - The High Impact Of The CPU
Most of you will certainly now it, the Voodoo2 offers 3D performance far above everything you've seen so far in the PC 3D gaming market. At this time it's absolutely save to say that Voodoo2 is faster than any other 3D gaming solution on the market. However, how much faster it is and what it takes to get the most out of it is still something to be evaluated.
The Voodoo2 is pretty much the first ever 3D chip which shows you the limitations of your CPU in a severe fashion. Whilst in the past the 3D accelerator used to be the bottleneck in 3D gaming performance, now with the Voodoo2 the bottleneck is the CPU in most of the cases. Testing shows that it takes at least a Pentium II 266 to see a significant impact of the Voodoo2 performance in game benchmarks and this only in case of high resolutions. At 640x480 most games won't let the Voodoo2 show it's full performance even in a Pentium II 300 system, which means that most likely even the upcoming Pentium II 400 will maybe only just deliver enough CPU power to max out the Voodoo2 . The K6 3D is the only Socket 7 candidate that could take the Voodoo2 closer to its limits, no currently available Socket 7 CPU is able to supply enough CPU power to really use the vast 3D force of the Voodoo2 .
You may be wondering what I am going on about, so I will try explaining this issue to you. As you could already read in my article 3D Accelerator Benchmarks - CPU Scaling of 3D Graphic Chips , there are two important things that determine the 3D performance of a system. First the CPU has to calculate the basic 3D geometry of a game scene, which is a procedure that requires a lot of floating point calculations. This is why systems with 6x86MX CPUs are having a some problems with 3D gaming, while offering an amazing integer performance, the floating point performance is pretty low. This geometry data is sent to the 3D accelerator which now does the rendering. This can be done in several steps, triangle setup, one or more rendering passes ... As you old see in above mentioned article, you can get to the state where the CPU can supply the geometry data to the 3D accelerator as fast as it wants, the 3 accelerator is just only able to process up to a certain amount of that data and the CPU has to wait until the 3D accelerator is finished before sending the next scene data. In this case you will not experience much or any increase in frame rate even when using a much faster CPU.
As you can imagine, there's also the possibility that it could be the other way around. the Quantum3D Obsidian 100SB with it's vast 3D performance based on up to 6 3Dfx Voodoo chips (instead of 2 in the normal Voodoo boards) already showed this phenomenon, the Voodoo2 cards show it even more severely. In this case the 3D accelerator is waiting for the CPU to deliver the geometry data. This means that e.g. overclocking the 3D accelerator chip will not improve 3D performance at all, because the 3D accelerator chip is simply waiting more clock cycles for the 3d geometry data from the CPU. You can also see the surprising effect that frame rate would not drop at all when switching to higher resolutions and adding another Voodoo2 board for SLI will not increase frame rate either, but offer you an even higher resolution at the same frame rate as the lowest resolution.
In the case of the Voodoo2 you will require at least a Pentium II 266 to start seeing a significant difference in frame rate between 640x480 and 800x600 frame rate, in case of two boards in SLI configuration a Pentium II 333 isn't even fast enough showing a frame rate difference between 640x480 and 800x600. This shows that the frame rate could be a lot higher at 640x480 if the Voodoo2 would get enough data from the CPU. 800x600 and more is occupying the Voodoo2 more, so the time of waiting for the CPU is less.
This is of course not the case for all games. It depends on how many calculations the CPU has got to do before it can send the data to the 3D chip. Actual games that do not require that much CPU power are Acclaim's Forsaken and Id's good old GLQuake. In case of GLQuake it only really shows in the pretty obsolete 'Timedemo demon ' benchmark, if you use heavy multiplayer benchmark demos like 'bigass1' the story is already different and this is most likely the case for Forsaken in network multi player mode as well. Games as Rage's Incoming or Id's Quake II are needing a lot of CPU power, so that above said is particularly valid for these games. The trend is going to more complex games though, which means that the effects of the CPU on Voodoo2 's frame rate will most likely get more rather than less, something you should consider when putting together a kick ass 3D gaming platform.
I'd also like to mention AGP in this context. AGP is supposed to offer us better or faster 3D experience. If you look at it closely though, you will see that this isn't the case at all currently. If the CPU can't deliver data to the 3D accelerator fast enough it doesn't really matter what kind of bus system is used transferring the data to the 3D chip. PCI is absolutely fast enough for that. Even in case of a Pentium II 400 or higher it's very unlikely that you will see much of a difference in 3D performance when using AGP instead of PCI. AGP is only having an advantage in case of texture transport from main memory, but if the texture is already in the on board memory of the 3D accelerator it doesn't come into account and receiving the texture data is only one of many steps of the rendering process of a 3D accelerator. I pretty much doubt that Voodoo2 is impacted by too slow texture transport when used as PCI solution at all, which is why I think it's about time dropping the discussion about a Voodoo2 AGP solution.
Please have a look at the benchmark results which back up the above said.