This article was originally published in 2003 in three separate articles. The article was written by Joel Hruska, who was a contributing writer to Sudhian at the time.  The information has now been combined into one long article - resource if you will - on anything and everything you ever wanted to know about the greatest video card that never was. 

A Fallen Titan’s Final Glory, Part I:  Setting the Stage

The decline and collapse of 3dfx in the waning months of 2000 was an untimely end for the company that had virtually invented and popularized the consumer-level 3D accelerator.  Once considered a virtually-unassailable market giant, 3dfx’s fa硤e of invulnerability hid major stress points that, through a mixture of external competition and internal mistakes, ultimately fractured and brought the company down.  The collapse was so dramatic it caught the entire market by surprise; while 3dfx did appear to be in trouble, not many people thought the situation was as bad as the events of December 2000 showed them to be. 

The Voodoo5 6000 (for those six of you who don’t know) was 3dfx’s last major product, and, though it was never released, it was the last to be manufactured in any kind of volume.  From its inception it was an attention-grabber, though its mammoth physical design and external power source would attract as much criticism as praise throughout its short and unofficial lifetime.  Based on the VSA (Voodoo Scaling Architecture) 100, the V5 6K was the ultimate over-the-top product; a 128 meg card when 32 meg was standard and 64 extravagant, a 4 chip architecture when other cards used only one (two at the most) and, inevitably, a (proposed) price tag that could choke a donkey.

It’s not possible to discuss the V5 6000 (and the entire VSA-100 product series) without also discussing the market conditions and historical events (can events from seven years ago BE historical?) that led to the ultimate demise of 3dfx.  Pay attention?there will be a quiz later.

The Dawning of a New Era:  1996-1997

Prior to 1996 and the introduction of the 3dfx Voodoo architecture, the world of computer gaming was a vastly different place.  Although primitive 3D environments were used in some games, even the fastest modern processors were unable to render these environments in high resolutions (for the time) while maintaining acceptable frame rates.  Even Quake, which was id’s first 3D game that allowed for jumping, climbing, and point-of-view rotation, couldn’t be run in a basic 640x480 resolution without looking like a slideshow.

This was a situation that benefited CPU giant Intel immensely.  With game performance dependent on processor performance, Intel could count on a steady crop of gamers buying or upgrading CPU’s, especially as games continued to demand more and more power in order to run smoothly.  Computers that’d whizzed by on Doom only a few years earlier choked on Quake; locking gamers into an upgrade cycle that distinctly benefited Santa Clara?until 3dfx introduced the original Voodoo architecture and torpedoed Intel’s carefully laid plans.

The original MMX instruction set (which boosted multimedia and video performance) was partially designed to increase game performance (and further boost sales).  By providing CPU’s with support for instruction sets that would increase 3D and multimedia performance Intel was both distancing itself from competitors like Cyrix, IDT, and AMD by offering features these companies did not possess, and introducing a feature set for which it could charge additional premiums.  MMX might’ve been only the first in a long line of instruction sets designed to boost 3D performance (and Intel profits)?if 3dfx hadn’t come along.

The Voodoo wasn’t the first 3D accelerator on the market, but it was the first 3D card on the market worth using.  The S3 Virge, in fact, performed so badly that it earned itself the nickname 3D "Decelerator" while Rendition’s Verite technology offered competitive video performance to a bowl of lukewarm tapioca?some of the time.  The Voodoo, in contrast, was fast, powerful, packed a whopping 4 meg of RAM on board, and was offered as an add-on card to a user’s primary video adapter.  Since many users invested in high-end 2D video cards at that time, releasing the Voodoo as a 3D add-in card actually made a great deal of sense, and didn’t force 3dfx to incur additional expenses by building a capable 2D engine or force users to choose between top-of-the-line 2D and gaming 3D performance.

For readers wondering just how much of a performance difference an original Voodoo could make we’ve included numbers gleaned from older sites where data is still available on GLQuake vs. ordinary Quake performance.  A Pentium 200 MMX system averaging a frame rate of 41 fps in Timedemo1 in Quake 1 (run in 320x200) rockets to 70 fps in GLQuake when a Voodoo card is used, for a whopping 70% boost in performance.  In 512x384 (probably the highest marginally playable resolution for this time period) we see Quake running at 21.2 fps on the Pentium 200 in software mode.  When we introduce the Voodoo we see performance jump no less then 300%, topping out an amazing 62 fps. 

Given these kind of results it’s easy to see why Voodoo cards sold hand over fist?and why Intel would have reason to be less than happy about these developments.  Although CPU power would continue to be an important factor in 3D performance, the emergence of the Voodoo knocked it firmly into second place when it came to determining just how fast a system would run in a given game. 

1996 and 1997 saw the mass adoption of the Voodoo, the introduction of 3dfx’s proprietary 3D API Glide (which came very close to dominating the market) and the establishment of 3dfx as *the* premiere 3D accelerator.  Most of 3dfx’s competitors (Matrox, ATI, S3, and Rendition) hadn’t really fielded seriously competitive products.  NVIDIA’s RIVA128 came on the scene in the fall of 1997 and became a popular low-cost option for OEMs, though it failed to make much of a dent in the higher-cost video market.

In February of 1998 NVIDIA launched the RIVA 128ZX, an upgraded version of the original with an 8 meg frame buffer.  The card took two slot spaces to install, was exceptionally loud, and offered only mediocre performance compared to its?.whoops.  Wrong card.  The RIVA 128ZX didn’t have many improvements beyond a larger frame buffer, but NVIDIA did have a full OpenGL ICD ready to go at launch.

3dfx did attempt to go after the 2D / 3D market (which would’ve put them in more-direct competition with NVIDIA), but their product of choice was the ill-fated Voodoo Rush.  Based on an out-sourced 2D design coupled with 3dfx Voodoo technology the card was plagued with compatibility problems and ultimately failed to have much market impact. 

Rampage = MIA

The Problem Lurking Behind Napalm:  Rampage = MIA.

The scenario I've laid out above is probably the most optimistic scenario 3dfx could've expected, even assuming perfect roll-outs and deadlines.  I've further assumed that the company might've made some decisions differently, pursued DDR rather than SDR-based solutions, and would've had little difficulty transitioning to a .22 or .18 micron-based solution. 

Even if all of this had occurred, however, 3dfx would still have had a problem.  As of December 2000, Rampage was only just taping out and NVIDIA's GF3 was set to launch in mere weeks.  Purists may claim that it was several months more before GF3 cards appeared (and that games actually utilizing the features didn't show up until nearly a year after that), but as both 3dfx and NVIDIA found, selling features (even if those features aren't available in modern software) is the key to selling video cards. Performance-wise, a hypothetical Voodoo5 6500 would've demolished a GeForce3 (at least in all modern games), but the card's DX6 / 7 feature-set would've appeared antiquated next to the GF3's programmable pixel and vertex shaders.  Stuck with an older architecture and no clear launch date on its next-generation technology, 2001 might've looked like a repeat of 1999 / 2000 for 3dfx.  A transcendent, powerful Napalm technology would've helped 3dfx but would not, by itself, have saved the company.

What was Rampage?

Rampage was the code-name for 3dfx's next-generation technology.  The radical new feature of the architecture (in addition to its DDR memory and .18 micron architecture) was its external T&L unit, code-named Sage.  Given their focus on parallel chip architectures, 3dfx had a choice to make?either design a chip with a dedicated hardware T&L unit for each CPU, or design a single external T&L unit that was powerful enough to serve all the GPU's simultaneously.  3dfx chose the latter approach, designing Sage to fuel the necessary T&L calculations to whichever GPU needed them at the moment.

We'll never know how successful this approach would've been, but one immediate problem that would've had to be dealt with is the latency involved in transmitting the calculations to and from the external T&L unit, as well as the need to carefully load balance the unit itself.  External designs are typically integrated later in a product's life (the cache integration that took place from the P2 to the Coppermine P3 or from K7 to Thunderbird being good examples of this).  However 3dfx was working on card designs that would allow each GPU access to its own Sage unit.  Eschewing the massive 4 GPU Voodoo5 6000 design, 3dfx had no plans to expand Rampage beyond a 2 GPU configuration (again, each potentially with its own T&L unit). 

Other features in Rampage included the M-buffer (a second-generation T-buffer unit) which would've supported all the features its predecessor did as well as adding other (unspecified) improvements.  Supposedly the M-buffer would've allowed for 2x and 4x FSAA without any type of performance hit.  Whether or not this is actually or was marketing hype is open to debate, however, for 3dfx is on record as having claimed this for T-buffer as well.  As appealing as the "something-for-nothing" mentality is, its rare-to-never that you find such scenarios in computing.  Even today NVIDIA and ATI haven't yet found a way to implement FSAA without taking some type of performance hit (though admittedly, if you're running in low resolutions the hit may be very small).

Was this a company-saving design?  Hard to tell.  It would've added AGP 4X support as well as anisotropic filtering (128 tap maximum, used to correct the fact that the M-Buffer lacked a texture clean-up filter).  It also would've included the first hardware support gained from 3dfx's purchase of GigaPixel, which supposedly would've eliminated up to 30-50% of overdraw with only a 1% rate of visual error.

Hidden in the above sentence, however, may be part of the reason the Rampage ultimately never saw the light of day.  First intended as a follow-up to the Voodoo2 (according to some reports we've seen) it was pushed back again?.and again (by Banshee)?then re-designed?then delayed, as Napalm became a stop-gap in its stead?and then, even as late as 2000, we see 3dfx planning new features for the core.

Delaying a core while you constantly attempt to stuff new features into it doesn't work.  It may be that 3dfx's visionary ideas, while commendable, ultimately caused delay after delay in Rampage deployment until the card barely taped out before disaster struck.

< previous - (18 of 21) - next >

---

• Setting the Stage (1 of 21)
• The Golden Era (2 of 21)
• The Throne Trembles (3 of 21)
• The Storm Breaks (4 of 21)
• A Titan Falls (5 of 21)
• The Voodoo 5 6000 (6 of 21)
• Our Card (7 of 21)
• Test Bed and Drivers (8 of 21)
• Benchmarking Methodology (9 of 21)
• 3dMark 2000 (10 of 21)
• 3dMark 2001 (11 of 21)
• Evolva (12 of 21)
• Expendable (13 of 21)
• D3D Performance (14 of 21)
• Quake 3 (15 of 21)
• Jedi Knight & Serious Sam (16 of 21)
• Analyzing the Numbers (17 of 21)
• Rampage = MIA (18 of 21)
• Did Banshee kill 3dfx? (19 of 21)
• Conspiracy Theories, cont. (20 of 21)
• The lasting impact of 3dfx (21 of 21)