CPU Scaling Of 3D Graphic Chips
After having a close look at the benchmark results of the different graphic cards we learned that you cannot predict the 3D performance of a 3D accelerator in your own system by just transferring the 3D performance results of a system that's faster or slower than yours. This means that if you only get 3D performance results of a high end Pentium II system, you may think that the slowest card in this comparison is also the slowest card in your K6 system, which we now know is seriously wrong.
The reason to this strange occurrence is that 3D chips don't necessarily scale linearly with CPU performance. A 3D graphic chip has its very own maximal polygon and pixel fill rate, which simply limits its 3D performance. On the other side, a CPU has got its own maximal 3D geometry setup performance, mainly determined by its FPU, which limits the amount of geometry data it can send to the graphic chip. One of these two determine the frame rate scored in a 3D benchmark. If the CPU is lacking of 3D geometry setup performance (pretty much equal FPU performance), the graphic chip has to wait for geometry data from the CPU, so that the frame rate is determined by the CPU performance. However, if the graphic chip is at its limit of pixel fill rate or polygon rate, the CPU can try sending geometry data to the graphic chip as fast as it wants, the graphic chip can't produce a higher frame rate anymore. In this case the frame rate is not determined by the CPU at all, it's the graphic chip that limits it.
Now what does this mean to us?
After publishing my first 3D benchmark results with a Pentium II 300 and a 6x86MX PR200, I received a very rude mail from a guy who accused me of lying, because in his Pentium MMX 166 system, his Voodoo Rush card scored only about 1 fps less than what I had published as the result of the same card running in a Pentium II 300 system. He could not believe that with a CPU which is 3 to 4 times as fast as his, the frame rate would only increase by a few percent. His argumentation was certainly sensible, but he forgot that it could be, that the Voodoo Rush chip is either scaling very slowly or maybe already at its performance limit in a small Pentium MMX 166 system.
The same wrong estimation could happen if you look at how the Apocalypse 3Dx scored in the system with the Pentium II 300. It reaches 5th place, scoring a very decent 77% of Voodoo performance. However, if you have got a system with e.g. a quite respectable 6x86MX 200, you should not expect the PowerVR chip to score 5th in this system as well. In this system it drops down to 12th place, only offering 58% of Voodoo performance anymore.
If you want to buy a 3D graphic card, you got to realize the 3D geometry performance of your system. Currently the IDT WinChip is offering the worst, the next one is the 6x86MX, then the K6, followed by the Pentium MMX and the best 3D geometry performance in X86 systems you'll of course find in Pentium Pro/Pentium II systems. Then you have to know if you want to play Quake, Quake II or any other game that's using this engine. The Quake/Quake II engine is the most FPU power hungry 3D engine, Turok e.g. is not as particular about the FPU performance and many other Direct3D games like e.g. Forsaken or Incoming aren't either. After you are aware of these two parameters, you want to take a look at the 'CPU Scaling of 3D Chips' charts below.
Now if you have got a powerful 3D system with a Pentium MMX or Pentium II, you want to go for a 3D chip that reaches high absolute results in frame rate as well as scales well over the CPUs. The number one chip to achieve this is currently the RIVA 128 from NVidia, which obviously is far from its maximal 3D performance even in a system with Pentium II 300. Videologic's PowerVR chip is another one that scales well, but doesn't score as high, and the same is valid for 3DLabs' Permedia 2. 3Dfx Voodoo only scales excellent if you buy it in form of a Quantum3D Obsidian 100SB, the normal Voodoo cards reach their absolute 3D performance limit somewhere between Pentium MMX 233 and Pentium II 300. However, the absolute 3D performance of the Voodoo is what makes it still the best choice for a high end system as well. The soon released Voodoo 2 will solve this problem and scale much better.
The owner of a mid range to low end system still wants a 3D chip that scores high absolute, but it doesn't matter if the chip scales well, as long as the performance is good in systems with lower 3D geometry performance of the CPU. The number one choices in this case are besides the 3Dfx Voodoo, which always is an excellent choice, the 3Dfx Voodoo Rush and the Rendition Verite 2x00. In my benchmark the Voodoo Rush chip is the one that scales least. It increases the Turok frame rate by unbelievable 11% from a system with a 6x86MX PR200 to a Pentium II 300. The Verite 2x00 scales better, but it reaches its limit somewhere around the Pentium MMX 233 and the Pentium II 233.
OK, so far about the theory, have a look at the results yourself.