3Dfx is a San Jose based manufacturer of 3D graphics accelerator hardware for arcade games, game consoles, and PC boards. Their official website is www.3dfx.com. 3Dfx does not sell any boards, but other companies do, e.g. Quantum3D.
Quantum3D started as a 3Dfx spin-off, manufacturing high end accelerator boards based on 3Dfx chip technology for consumer and business market, and supplying arcade game technology. See their home page at www.quantum3d.com for additional information. For general inquiries regarding Quantum3D, please send mail to info@quantum3d.
The Voodoo Graphics (tm) is a chipset manufactured by 3Dfx. It is used in hardware acceleration boards for the PC. See the HOWTO section on supported hardware.
The Voodoo Rush (tm) is a derivate of the Voodoo Graphics (tm) that has an interface to cooperate with a 2D VGA video accelerator, effectively supporting accelerated graphics in windows. This combo is currently not supported with Linux.
The Voodoo 2 (tm) is the successor of the Voodoo Graphics (tm) chipset, featuring several improvements. It is announced for late March 1998, and annoucements of Voodoo 2 (tm) based boards have been published e.g. by Quantum 3D, by Creative Labs, Orchid Technologies, and Diamond Multimedia.
The Voodoo 2 (tm) is supposed to be backwards compatible. However, a new version of Glide will have to be ported to Linux.
The Voodoo Graphics (tm) (but not the Voodoo Rush (tm)) boards are add-on boards, meant to be used with a regular 2D VGA video accelerator board. In short, the video output of your regular VGA board is used as input for the Voodoo Graphics (tm) based add-on board, which by default passes it through to the display also connected to the Voodoo Graphics (tm) board. If the Voodoo Graphics (tm) is used (e.g. by a game), it will disconnect the VGA input signal, switch the display to a 640x480 fullscreen mode with the refresh rate configured by SST variables and the application/driver, and generate the video signal itself. The VGA doesn't need to be aware of this, and won't be.
This setup has several advantages: free choice of 2D VGA board, which is an issue with Linux, as XFree86 drivers aren't available for all chipsets and revisions, and a cost effective migration path to accelerated 3D graphics. It also has several disadvantages: an application using the Voodoo Graphics (tm) might not re-enable video output when crashing, and regular VGA video signal deteoriates in the the pass-through process.
Voodoo Graphics (tm) chipsets have two units. The first one interfaces the texture memory on the board, does the texture mapping, and ultimately generates the input for the second unit that interfaces the framebuffer. This one is called Texelfx, aka Texture Management Unit, aka TMU. The neat thing about this is that a board can use two Texelfx instead of only one, like some of the Quantum3D Obsidian boards did, effectively doubling the processing power in some cases, depending on the application.
As each Texelfx can address 4MB texture memory, a dual Texelfx setup has an effective texture cache of up to 8MB. This can be true even if only one Texelfx is actually needed by a particular application, as textures can be distributed to both Texelfx, which are used depending on the requested texture. Both Texelfx are used together to perform certain operations as trilinear filtering and illumination texture/lightmap passes (e.g. in glQuake) in a single pass instead of the two passes that are required with only one Texelfx. To actually exploit the theoretically available speedup and cache size increase, a Glide application has to use both Texelfx properly.
The two Texelfx can not be used separately to each draw a textured triangle at the same time. A triangle is always drawn using whatever the current setup is, which can be to use both Texelfx for a single pass operation combining two textures, or one Texelfx for only a single texture. Each Texelfx can only access its own memory.
Voodoo Graphics (tm) chipsets have two units. The second one interfaces the framebuffer and ultimately generates the depth buffer and pixel color updates. This one is called Pixelfx. The neat thing here is that two Pixelfx units can cooperate in SLI mode, like with some of the Quantum3D Obsidian boards, effectively doubling the frame rate.
SLI means "Scanline Interleave". In this mode, two Pixelfx are connected and render in alternate turns, one handling odd, the other handling even scanlines of the actual output. Inthis mode, each Pixelfx stores only half of the image and half of the depth buffer data in its own local framebuffer, effectively doubling the number of pixels.
The Pixelfx in question can be on the same board, or on two boards properly connected. Some Quantum3D Obsidian boards support SLI with Voodoo Graphics (tm).
As two cards can decode the same PCI addresses and receive the same data, there is not necessarily additional bus bandwidth required by SLI. On the other hand, texture data will have to be replicated on both boards, thus the amount of texture memory effectively stays the same.
There are now two types of Quantum3D SLI boards. The intial setup used two boards, two PCI slots, and an interconnect (e.g. the Obsidian 100-4440). The later revision which performs identically is contained on one full-length PCI board (e.g. Obsidian 100-4440SB). Thus a single board SLI solution is possible, and has been done.
The most essential difference between different boards using the Voodoo Graphics (tm) chipset is the amount and organization of memory. Quantum3D used a three digit scheme to descibe boards. Here is a slightly modifed one (anticipating Voodoo 2 (tm)). Note that if you use more than one Texelfx, they need the same amount of texture cache memory each, and if you combine two Pixelfx, each needs the same amount of frame buffer memory.
"SLI / Pixelfx / Texelfx1 / Texelfx2 "
1/2/2/0 solution, with the minimally
required total 4Mb of memory. A Canopus Pure 3D
would be 1/2/4/0, or 6MB. An Obsidian-2220
board with two Texelfx would be 1/2/2/2,
and an Obsidian SLI-2440 board would be 2/2/4/4.
A fully featured dual board solution (2 Pixelfx, each
with 2 Texelfx and 4MB frame buffer, each Texelfx 4 MB
texture cache) would be 2/4/4/4, and the
total amount of memory would be
SLI*(Pixelfx+Texelfx1+Texelfx2), or 24 MB.
So there.
No. The Voodoo Graphics (tm) architecture uses 16bpp internally. This is true for Voodoo Graphics (tm), Voodoo Rush (tm) and Voodoo 2 (tm) alike. Quantum3D claims to implement 22-bpp effective color depth with an enhanced 16-bpp frame buffer, though.
No. The Voodoo Graphics (tm) architecture uses 16bpp internally for the depth buffer, too. This again is true for Voodoo Graphics (tm), Voodoo Rush (tm) and Voodoo 2 (tm) alike. Again, Quantum3D claims that using the floating point 16-bits per pixel (bpp) depth buffering provides 22-bpp effective Z-buffer precision.
The Voodoo Graphics (tm) chipset supports up to 4 MB frame buffer memory. Presuming double buffering and a depth buffer, a 2MB framebuffer will support a resolution of 640x480. With 4 MB frame buffer, 800x600 is possible.
Unfortunately 960x720 is not supported. The Voodoo Graphics (tm) chipset requires that the amount of memory for a particular resolution must be such that the vertical and horizontal resolutions must be evenly divisible by 32. The video refresh controller, though can output any particular resolution, but the "virtual" size required for the memory footprint must be in dimensions evenly divisible by 32. So, 960x720 actually requires 960x736 amount of memory, and 960x736x2x3 = 4.04MBytes.
However, using two boards with SLI, or a dual Pixelfx SLI board means that each framebuffer will only have to store half of the image. Thus 2 times 4 MB in SLI mode are good up to 1024x768, which is the maximum because of the overall hardware design. You will be able to do 1024x768 tripled buffered with Z, but you will not be able to do e.g. 1280x960 with double buffering.
Note that triple buffering (no VSync synchonization required by the application), stereo buffering (for interfacing LCD shutters) and other more demanding setups will severely decrease the available resolution.
The maximum texture size for the Voodoo Graphics (tm) chipset is 256x256, and you have to use powers of two. Note that for really small textures (e.g. 16x16) you are better off merging them into a large texture, and adjusting your effective texture coordinates appropriately.
The Voodoo Graphics (tm) hardware and Glide support the palette
extension to OpenGL. The most recent version of
Mesa does support the
GL_EXT_paletted_texture
and
GL_EXT_shared_texture_palette extensions.
If you want to put aside considerations about warranty and overheating, and want to do overclocking to boost up performance even further, there is related info out on the web. The basic mechanism is to use Glide environment variables to adjust the clock.
Note that the actual recommended clock is board dependend. While the default clock speed is 50 Mhz, the Diamond Monster 3D property sheet lets you set up a clock of 57 MHz. It all comes down to the design of a specific board, and which components are used with the Voodoo Graphics (tm) chipset - most notably access speed of the RAM in question. If you exceed the limits of your hardware, rendering artifacts will occur to say the least. Reportedly, 57 MHz usually works, while 60 MHz or more is already pushing it.
Increasing the clock frequency also means increasing the waste heat disposed in the chips, in a nonlinear dependency (10% increase in frequency means a lot larger increase in heating). In consequence, for permanent overclocking you might want to educate yourself about ways to add cooling fans to the board in a way that does not affect warranty. A very recommendable source is the "3Dfx Voodoo Heat Report" by Eric van Ballegoie, available on the web.
There is a FAQ by 3Dfx, which should be available at their web site. You will find retail information at the following locations: www.3dfx.com and www.quantum3d.com.
Inofficial sites that have good info are "Voodoo Extreme" at www.ve3d.com, and "Operation 3Dfx" at www.ve3d.com.