I first want to say how rewarding it is to read all your reviews after having worked on the design of Voodoo Graphics (the chipset on the Orchid Righteous 3D board) for over two years. I am one of the founders of 3Dfx and one of our goals was to deliver the highest quality graphics possible to the PC gamer. It was and still is a very risky proposition because of the cost sensitivity of the marketplace. But your reviews help convince me that we did the right thing.

I thought I would share with you a little bit about what is inside the 3Dfx Voodoo Graphics chipset. There are 2 chips on the graphics board. Each is a custom designed ASIC containing approximately 1 million transistors. Although this number of transistors is on the order of a 486, it is a lot more powerful. Why? Because the logic is dedicated to graphics and there’s a lot of logic to boot. For example, bilinear filtering of texture maps requires reading four 16-bit texels per pixel (that’s 400 Mbytes/sec at 50 Mpixels/sec) and then computing the equation red_result=r0*w0+r1*w1+r2*w2+r3*w3 where r0:3 are the four red values and w0:3 are the four weights based on the where the pixel center lies with respect to the four texels. This is performed for each color channel (red, green, blue, alpha) resulting in 16 multiples and 12 additions or 28 operations per pixel. At 50 Mpixels per second that is 1,400 Mops/sec. The way this is designed in hardware is you literally place 16 multipliers and 12 adders on the chip and hook them together. And this is only a small part of one chip. There are literally dozens of multipliers and dozens of adders on each of the two chips dedicated only to graphics. Each chip performs around 4,000 million actual operations per second, of which around one third are integer multiplies. These are real operations performed - if you were to try to do these on a CPU (or a DSP) you must also do things like load/store instructions and conditions. In my estimation it would take about a 10,000 Mip computer (peak) to do the same thing that one of our chips does. This is about 20 of the fastest P5-200 or P6-200 chips per one of our chips. Not exactly cost-effective. So if you want to brag, you can say your graphics card has approximately the same compute power as 40 P5-200 chips. Of course, these numbers are more fun than they are meaningful. What is meaningful in graphics is what you see on the screen.

Now of course, if you were writing a software renderer for a game, you wouldn’t attempt to perform the same calculations we perform on our chip on a general purpose CPU. You would take shortcuts, like using 8-bit color with lookup tables for blending, or performing perspective correction every ‘n’ pixels. The image quality will depend on how many shortcuts you take and how clever you are. Voodoo Graphics takes no shortcuts and was designed to give you the highest quality image possible within the constraint of 2 chips. As your reviews have shown, it is evident that you can see the difference in quality and performance.

Now I am sure the subject of triangle setup and geometry calculations will come up sooner or later in this newsgroup. Let me make a preemptive strike and answer your questions before you ask them. There is no geometry acceleration on the board, where geometry is defined as geometric transformation and lighting. In the Wizard’s tower demo you see lighting being applied to texture maps through the use of a ‘lighting map’, which is another texture map that contains the results of off-line radiosity calculations. This is not traditional lighting in the OpenGL sense, but is nonetheless a very powerful method of performing static lighting. It is becoming more popular with games and personally, I think its great! It requires bilinear filtering AND high fill rates, both of which the R3D card has.

Now back to triangle setup. The Voodoo Graphics chipset performs about 2/3 of triangle setup in hardware. When designing Voodoo Graphics I carefully studied exactly what triangle setup our design required and we placed things that were hard for a Pentium to perform in hardware and left things easy for a Pentium to perform out of the hardware design. Our triangle engine is also very efficient in that it requires less setup than most (I worked 9 years at SGI and have a lot of triangle engine experience). The net result is that the 1/3 of triangle setup we perform on the Pentium is not many cycles at all. That is why our triangle numbers are so high. With an efficient design, you can afford to use the Pentium to perform some of your triangle setup. With an inefficient design, you cannot. I know this is a very controversial subject, so I will stop right here.

I hope this answers some of your questions in advance. Thanks for buying the board and I hope you enjoy it. As for the flight sims, I am waiting for one too, and am anxiously awaiting to see what our chips can do. I wrote the original SGI flight simulator and hopefully, I won’t have to write another one :-}