Missing in action: Whereabouts now known - Feature #29

Once high tech, now low profile. James Wang hits the road to track down the companies we all loved but have since dropped off the radar.

3dfx Interactive
3dfx brought 3D to the PC. With insanely successful cards (Voodoo 1 and 2), a proprietary API (Glide) and the support of the entire industry, this company could have monopolised the market for years to come. In reality, it was lead by a CEO with a record of rehashing old products, the company itself failed to capitalise on its advantage, and made mistake after mistake, both managerial and technical. The internal mismanagement translated to future products (Voodoo 4 and 5) whose designs were both technically misdirected and hopelessly late. In fact, nearly every 3dfx product was late in one form or another.

The original Voodoo missed internal milestones and only launch thanks to a sudden drop in memory prices. The Voodoo Banshee was nine months late, the Voodoo 5 almost a year and the mythical 'Rampage' chip completed only after a total redesign and nearly four years of R&D resources. In the dieing days of December 2000, what would seem to the outsider like a fairy tale story ended with the selling of assets to NVIDIA. Soon after, the company's doors were closed and the company dissolved. More than a hundred engineers, including Chief Scientist Gary Tarolli joined NVIDIA where they met with old friends (Brian Burke, ex-3dfx PR manager), many of whom jumped the boat long before the bitter black Friday finale. The co-founder and Chief Technical Officer, Scott D.

Sellers, to this day remains somewhat embittered over the loss of the company and has maintained a low public profile.

Verdict: Horrific management; feature creep causing no execution; poor OEM business.
Contribution: Bringing true 3D acceleration to the PC.

Bitboys
Bitboys is a cursed company. Started in 1991 by ex-Future Crew members who were tired of the lack of graphical power on the PC, the company set out to design its own chips. Its first chip, the Pyramid3D, was developed in collaboration with fab partner Tritech. When the chip finally made it to silicon, Tritech lost a major law suit on an audio patent infringement case. With auditors banging on doors for royalties, Tritech went bust and along with it, Bitboys' first graphics processor. The Pyramid3D, by 1997 standards, (Voodoo 1) is almost a fantasy. It featured programmable pixel shaders that would be the rough equivalent of our DirectX 8 standard today, as well as a geometry engine -- a term unheard of until the Geforce256. This engine featured a fully programmable pipeline that supported loops and subroutines. The degree of flexibility in certain areas comes close to today's latest R300 and NV30 GPUs.

With no partner to produce their chip, Bitboys continued to design its next processor, which places heavy emphasis on bandwidth through the use of EDRAM. Long story cut short: after several revisions and troubles with its new partner Infineon, the 'Avalanche3D' GPU reached working condition. As if fate was against them, in the same quarter, Infineon posted a huge loss and decided to shut down its EDRAM-based production facilities. Bitboys once again was stranded with a finished product and no fab partner. Faced with financial difficulties, its Dallas office in the U.S. was shut and it looked to greener pastures. In just a few weeks, its engineers whipped up a vector-based processor for PDAs and cell phones. Finally last year, Bitboys signed a deal for its first successful chip. It's unknown who the buyer is but most likely a Japanese mobile phone company. A new office was setup in Noormarkku after this success and Bitboys earned a cool EU150,000 profit for the last quarter of 2002. Good on it!
 
Case: Great dreamers; feature creep; bad luck, but persistence.
Contribution: Environment Mapped Bump Mapping (EMBM)

PowerVR
A new architecture is always refreshing, and PowerVR is the only company to have brought an alternative to immediate-mode rendering for PC graphics. An old player in the market, PowerVR competed along the sides of 3dfx, Rendition and Matrox in the early days of 3D. Its success peaked with the introduction of the KRYO II, based on the PVR Series3 design. This unexpected chip took market share in the retail channels thanks to a partnership with Hercules. Unfortunately the follow-up product never arrived.

STMicroelectronics, the board maker for KYRO-based products, decided to sell off its graphics division last February. In light of this loss, work on the Series4-based products was cancelled and Series5 development gained momentum. This time with DirectX 9 taking precedence, tile-based hidden surface removal will offer a different kind of advantage. As long shader instructions can take many cycles to complete, an early kill method like TBR can save enormous processing time.
 
Series5, due in 2003, will be based on a 0.13-micron process and is expected to have some degree of DirectX 9 compliance. PowerVR has yet to decide whether they will produce the chip with a new partner or employ foundries such as TSMC.

Case: Uncommitted to high-end 3D graphics; loss of fab partner.
Contribution: Proving an alternate rendering method exists for 3D graphics.

Matrox
Parhelia was the last straw for Matrox and it didn't come close to saving the company. Even their 2D quality reputation has been scarred due to buggy silicon that produced rare banding problems. Unfortunately, Parhelia is likely to the last GPU we see from Matrox for quite some time. There are many people reporting that Matrox has sacked its entire Florida graphics division -- the core engineers behind Parhelia.

The design of this chip fell victim to the same feature creep, mismanagement and lack of tools that plagued 3dfx. Insiders claim that the lack of tools was the main reason for the delay. ATI and NVIDIA can afford complex simulators like the IKOS, which emulates an entire GPU via its Field Programmable Gate Array (FPGA). Every tier one graphics company use these to speed up chip design through early simulation. While Matrox engineers were faced with the daunting task of creating an 80 million transistor GPU, upper management refused request after request for the purchase of advanced simulation tools. This, together with the relatively i