3dfx Interactive was a company which specialized in the manufacturing of cutting-edge 3D graphics processing units and, later, graphics cards. After dominating the field for several years in the late 1990s, by the end of 2000 it underwent one of the most high-profile demises in the history of the PC industry. It was headquartered in San Jose, California until, on the verge of bankruptcy, its intellectual assets (and many employees) were acquired by its rival, NVIDIA Corporation.
Founded in 1994, 3dfx released its famous Voodoo Graphics chip in 1996. The company only manufactured the chips and some "reference boards", and initially did not sell any product to consumers; rather, it acted as an OEM supplier for graphics card companies, who designed, manufactured, marketed, and sold their own graphics cards which included the Voodoo.
After a fortuitous drop in EDO RAM prices due to the volatile DRAM market, Voodoo Graphics cards became feasible for the consumer PC market. The Voodoo was the first 3D accelerator that actually accelerated graphics performance; most systems of the era performed no better, if not worse, than software rendering engines.
The Voodoo 1, as the Voodoo Graphics would be later known, was notable for its lack of an onboard VGA controller. This meant a Voodoo equipped PC still required a separate VGA graphics card. The Voodoo 1 occupied a separate PCI slot and only kicked in when the host PC ran a 3D game that had been programmed to use the Voodoo. A passthrough VGA cable daisy-chained the VGA card to the Voodoo 1, and then out to the monitor. Although this was a cumbersome arrangement, hardcore PC gamers were willing to put up with it to gain what was (then) the ultimate in 3D graphics.
The Voodoo 1's main competition was PowerVR and Rendition. PowerVR produced a 3D add-on card, whereas Rendition offered an integrated (3D+VGA) single-chip solution. Neither competitor achieved the Voodoo 1's popularity among gamers and developers.
In order to ensure better performance, 3dfx developed the proprietary Glide API for game developers to use while writing their 3D games. Glide exposed the Voodoo's internal hardware to application programmers directly; it was essentially a small subset of OpenGL that could be implemented in hardware. This strategy differed from that of other 3D APIs of the era (Direct3D, OpenGL, and QuickDraw 3D), which all hid low-level hardware behind an "abstraction layer," with the goal of providing application developers a standard, hardware-neutral interface.
The ostensible advantage of an abstraction layer is that game developers save programming effort by writing their 3D rendering code one time, for a single API, and the abstraction layer takes care of managing all the differences in all the hardware cards available to consumers. This advantage is still in place today. However, in the early days of the 3D graphics card, Direct3D and OpenGL were less mature than today, and computers were much slower and had less memory. The abstraction layers' overhead crippled performance in practice. 3Dfx had therefore created a strong advantage for itself by aggressively promoting Glide, which was actually implemented in hardware, and therefore effectively eliminated the speed and memory problems of an abstraction layer. Although a full OpenGL library was available for the Voodoo, most developers were glad to spend extra programmer time to support Glide instead, so their games would run faster and look better.
The killer application for the Voodoo was the game Quake, by id Software. Quake's rendering code used a stripped-down version of OpenGL known as MiniGL, which was designed with the Voodoo in mind, and, predictably, ran particularly quickly on the Voodoo.
In August 1997, 3dfx released the Voodoo Rush chipset, combining a Voodoo chip with a 2D chip from Alliance Semiconductor that lay on the same circuit board, eliminating the need for a separate VGA card. Unfortunately it performed worse than the Voodoo 1, primarily owing to the fact that the 2D and 3D cores shared the same memory interface and couldn't master the PCI bus correctly, incurring a 10% performance hit. Later versions released by Hercules had 8MB of ram and a 10% higher clock speed to close the performance gap, but in the marketplace the damage had already been done. A rare, third version was produced which featured a Cirrus Logic 2D chip instead of the earlier model. This version fixed the PCI bus collisions and memory interface problems, but there was little interest from graphics board manufacturers and only a few hundred units (if even that) were produced before the Rush was finally discontinued in early 1998.
In 1998, 3dfx released Voodoo's successor, the popular Voodoo 2. The Voodoo 2 was architecturally similar, but the basic board configuration added a second texturing unit, allowing two textures to be drawn in a single pass. The Voodoo 2 also had a faster clock rate (90MHz), a wider memory bus (192-bit, compared to Voodoo's 128-bit), and support for larger amounts of memory (up to 8MB texture / 4MB frame buffer compared to the Voodoo's 4MB texture / 2MB frame buffer.) A single Voodoo 2 board could display a maximum resolution of 800×600 with higher quality textures.
A problem with the Voodoo 2 was the fact that it required three chips and a separate VGA graphics card, whereas new competing 3D products, such as the ATI Rage Pro, NVIDIA RIVA 128, and Rendition Verite 2200, were single-chip products.
The Voodoo 2 introduced Scan-Line Interleave (SLI) to the gaming market. In SLI mode, two Voodoo 2 boards were connected together, each drawing half the scanlines of the screen. For the price of a second Voodoo 2 board, users could essentially double their 3D throughput. A welcome side effect of SLI mode was an increase in the highest supported resolution, up to a then-impressive 1024×768.
The Voodoo 2 SLI scheme was considered the pinnacle of gaming performance at the time. SLI was not used in any subsequent 3Dfx board designs. Later, in 2004, SLI technology was reintroduced by nVidia (which now owns 3dfx's intellectual property) in their GeForce line. One may add a second card and double some aspects of 3D rendering.
Near the end of 1998, 3dfx released the Voodoo Banshee, which used a lower price to aim at a more mainstream consumer market.
A single-chip solution, the Banshee was basically a legacy VGA core and part of a Voodoo 2 (one Texture Mapping Unit), clocked slightly faster than the Voodoo 2. The Banshee's single-chip form factor dictated a 128-bit memory bus, like the first Voodoo. Performancewise, the Banshee was a mixed bag. In scenes which used multiple textures per polygon, the Voodoo 2 was substantially faster, due to the 2nd TMU. In scenes dominated by single-textured polygons, though, the Banshee would match (or even slightly exceed) the Voodoo 2. While it was not a hit on the scale of the Voodoo 1 or 2, the Banshee sold a respectable number of units.
In 1997, 3dfx was working with Sega to develop Sega's next video game console. The process involved two competing designs: a unit called "Katana" being developed in Japan using NEC and VideoLogic technology vs. the "Blackbelt", a system designed in America using a GPU from 3dfx. This deal had the potential to get 3dfx's foot in the home console door, provided the Blackbelt became the console that would become the Dreamcast. Unfortunately for 3Dfx, Sega chose the NEC solution. 3dfx sued Sega for breach of contract when the Katana was chosen, accusing Sega of starting the deal in bad faith to take 3Dfx technology, and eventually the case was settled out of court; but the failure of the Blackbelt was 3dfx's own doing.
When 3dfx declared its Initial Public Offering (IPO) in April 1997, it made the mistake of revealing every detail of the contract with Sega. By law, when a company files an IPO in the United States, it has to make public all details of its business and financial situation, but sensitive information can be kept secret, so long as it does not materially affect the company's statement of its financial position and outlook. Sega had been keeping the development of its next-gen console secret during this competition, and was outraged when 3dfx publicly laid out its deal with Sega over the new system in the IPO; Sega quickly quashed the Blackbelt and used the Katana as the model of the Dreamcast.
In early 1998, 3dfx embarked on its "Rampage" development project, which was to be a new graphics card that would take two years to develop, and would supposedly be several years ahead of the competition once it debuted. The company hired hardware and software teams in Austin, Texas to develop 2D and 3D Windows device drivers for Rampage in the summer of 1998. The hardware team in Austin inially focused on Rampage, but then worked on Transformation and Lighting (T&L) engines and on MPEG decoder technology. (Later, these technologies were part of the NVidia asset purchase in December 2000.)
Voodoo 3 and strategy shift
In mid-1999 the Voodoo 3 was released, which was at heart a dual-core Voodoo 2 with Banshee's 2D core. It was a compelling solution, since an SLI-configured Voodoo 2 took up three slots, including the 2D card. However, due to the Voodoo 3's design legacy, it lacked support for several technologies that its competitors, ATI Technologies, Matrox and NVIDIA, had since integrated, most notably 32-bit color support and textures greater than 256*256 in size.
3Dfx executed a major strategy change just prior to the launch of Voodoo 3 by purchasing STB Systems, which was one of the larger graphics card manufacturers at the time; the intent was for 3Dfx to start manufacturing, marketing, and selling its own graphics cards, rather than functioning only as an OEM supplier. This alienated 3Dfx's OEM customers, all of whom chose to switch, and source their 3D chips from other manufacturers, rather than do business with a company who was their direct competitor at retail.
This strategy change was one of the main contributors to 3dfx's downfall; the company did not sell any Voodoo 3, 4, or 5 chips to third party manufacturers. The company was also presumably distracted by the need to focus both on the retail market as well as the OEM market, selling cards to computer manufacturers. The latter was hard-won business, but provided a steady income to fund subsequent development. A signficant requirement of the OEM business was the ability to consistently produce new products on the 6 month product refresh cycle the computer manufacturers required; 3Dfx did not have the methodology nor the mindset to the focus on this business model. In the end, 3Dfx opted to focus on the retail business using its own manufactured and branded products. The Voodoo 3 sold relatively well, but disappointingly compared to the first two models.
Voodoo 4 and 5
The company's next (and as it would turn out, final) product was code-named Napalm. Originally, this was just a Voodoo 3 modified to support newer technologies and higher clock speeds, with performance estimated to be around the level of the NVidia TNT2. However, Napalm was delayed, and in the meantime NVidia brought out their landmark GeForce chip, which shifted even more of the computational work from the CPU to the graphics chip. Napalm would have been unable to compete with GeForce, so it was redesigned to support multiple chip configurations, like the Voodoo 2 had. The end-product was named VSA-100, which stood for Voodoo Scalable Architecture.
The two initial products were the Voodoo 4 4500 (single chip) and the Voodoo 5 5500 (dual chip), with a further two parts, the Voodoo 5 5000 (dual chip, but with a smaller frame buffer) and the Voodoo 5 6000 (quad chip) due to be launched later. But by the time the VSA-100 based cards made it to the market, the second-generation GeForce cards had arrived, which offered substantially better performance. By this point ATI had also released their Radeon line, which performed competitively with the GeForce 2 line. The only real advantage the Voodoo 5 5500 had over the GeForce 2 GTS or Radeon was that it had a better anti-aliasing implementation, and didn't lose as much performance when antialiasing was enabled. Voodoo 4 4500 was beaten in almost all areas by the Geforce 2 MX — a low-cost board sold mostly as an OEM part for computer manufacturers — and Radeon VE.
The Voodoo 5 6000 never got to the market, due to a severe bug resulting in data corruption on the AGP bus on certain boards, and was limited to AGP 2*, which would have prevented its use on the then-new Pentium 4 motherboards. Later tests proved that while the Voodoo 5 6000 would have been able to outperform the GeForce 2 GTS, it would've been outperformed by the GeForce 2 Ultra and the GeForce 3. The Voodoo 5 5000 never got launched either, as the smaller frame buffer didn't significantly reduce cost over the Voodoo 5 5500.
Voodoo 4 was as much of a disaster as Voodoo Rush, and while Voodoo 5's sales were respectable, they were nowhere near as good as 3Dfx needed. In late 2000, several of 3dfx's creditors decided to initiate bankruptcy proceedings. 3Dfx would have had virtually no chance of successfully contesting these proceedings, and instead opted to be bought by NVidia, ceasing to exist as a company. Most of the design team working on "Rampage" (the successor to the VSA-100 line) was transferred to the team working on what has since become the GeForce FX series.
Cause for Decline
3dfx's decline is a matter of debate. Some attribute it to 3dfx lavishly spending on its employees and they reported spent $30,000 just on company lunches and other perks a month, even up to the last two weeks before it went under.
3dfx's fall is most often attributed to managerial prioritizing of research and development. When Greg Ballard became CEO of 3dfx in 1997, analysts marked that as a turning point since Ballard was a marketing guru but he failed to understand R&D. Voodoo cards were typically highly expensive, and left the mid and low end of the market to ATI and NVidia. NVidia chose short development cycles, whereas 3Dfx pursued lengthy, ambitious development cycles, and NVidia and ATI cards eventually ended up with better overall performance, with Matrox holding the edge in image quality.
The "Rampage" project, which 3Dfx put much effort into but never was able to bring to market, is said to have been technologically several years ahead of the competition. It debuted in 3Dfx's labs in December of 2000, within weeks of the sale of 3Dfx's assets to NVidia. The Rampage design team was using a pioneering synthesis tool set which was still under development as the design proceeded.
In addition, the company continued to vacillate on its commitment to the delayed Rampage project versus the need for short-term retail products, such as the Napalm/VSA-100. Because Rampage was oft-delayed — it had been scheduled to show at the 1998 Comdex — 2D and 3D driver software was up and running when it hit the lab.
However, the impending release of Rampage was too little, too late. The deal to "wind down" the company was less than 2 weeks from closure at that point. The history of and participants in the 3Dfx/NVidia deal making can be read in the respective companies financial filings from that time period. The resolution of those arrangements (with respect to 3Dfx's creditors and its bankruptcy proceedings) was still being worked through the courts as of September 2005.
While some have speculated that shipping the "Rampage" might have saved 3Dfx, the fact remains that the company never mastered the new concept of relatively cheap, high-performance dies with integrated 2D acceleration, which was to become the de facto standard of PC graphics cards very soon. The success of "Rampage" would not have simply depended upon raw performance, but also the cost of manufacturing, very much reflected in retail prices. According to documents from late in 3Dfx's life, the "Rampage" core was evidently little more than a more powerful version of the VSA-100, with an entirely separate chip code-named "Sage" required for T&L and hardware shader operation. It remains unknown whether "Rampage" would have been a practical product, let alone enough to keep the company alive in the card industry.
|Voodoo 1||1 Geometry Unit, 1 Texturing unit (no VGA)||50||50||4MB or 6MB||PCI|
|Voodoo Rush v1||1 Geometry Unit, 1 Texturing Unit, 1 Alliance Semiconductor 2D processor||75||75||4MB or 6MB||PCI|
|Voodoo Rush v2||1 Geometry Unit, 1 Texturing Unit, 1 Cirrus Logic 2D processor||80||80||6MB||PCI|
|Voodoo 2 1000||1 Geometry Unit, 2 Texturing Units||90||90||8MB or 12MB||PCI|
|Banshee||Single-Chip (3D+VGA)||100||100||16MB||AGP 1x/PCI|
|Voodoo 3 1000||Single-Chip||125||125||8MB||AGP 2x/PCI|
|Voodoo 3 2000||Single-Chip||143||143||16MB||AGP 2x/PCI|
|Voodoo 3 3000||Single-Chip||166||166||16MB||AGP 2x/PCI|
|Voodoo 3 3500||Graphics processor, A/V processor||183||183||16MB||AGP 2x|
|Voodoo 4 4500||Single-Chip||166||166||32MB||AGP 2x/PCI|
|Voodoo 5 5000||Two Graphics processors||166||166||32MB*||PCI|
|Voodoo 5 5500||Two Graphics processors||166||166||64MB**||AGP 2x/PCI|
|Voodoo 5 6000||Four Graphics processors||166†||166†||128MB***||AGP 2x|
- 3D+VGA - products before the Banshee contain only a 3D core (no legacy/VGA)
- †The Voodoo5 6000 was originally intended to have a core and memory clock of 183MHz, but all of the prototypes running at 183MHz stopped working after a short while. The only still-working VooDoo5 6000s all run at 166MHz, and 3dfx had decided to drop the 183/183MHz idea anyway.
- *Shared by two processors; effectively 16MB VRAM. However, the Voodoo 5 5000 was never launched.
- **Shared by two processors; effectively 32MB VRAM.
- ***Shared by four processors; effectively 32MB VRAM.
- Firingsquad article about the demise of 3dfx
- Overview of the history of 3D graphics chips
- Performance analysis of the unreleased Voodoo 5 6000
- Another analysis of the unreleased quad chip Voodoo 5 6000
- 3dfx Archive
- 3dfx's Dream is Cast Aside