Inside The Demise of 3dfx September 26, 2002 Brandon Bell

Introduction

Page:: ( 1 / 6 )

We’ve chronicled the humble beginnings of industry titans ATI and NVIDIA in the past, but for today’s article we’re doing something a bit different. Rather than discuss the origins of a 3D company you’re familiar with, 3dfx, we were given the unique opportunity to learn more about what was going on within the company around the time of its sudden downfall. However, unlike previous industry articles we’ve published, this one comes straight from the horse’s mouth!

For obvious reasons our source would like to remain anonymous, but we’ve known him for quite awhile and can assure you that he is indeed legit. He will briefly go over the early days of 3dfx, before going into detail over each of the company’s products. From the original Voodoo Graphics chipset, all the way to unannounced parts such as Fearless and Mojo, it’s all covered here. So without further discussion, lets listen up to what he has to say!

In the beginning…

It was a sad loss for the entire graphics industry when 3dfx announced they were closing their doors. Within the last year and half there have been several articles on the subject of 3dfx's demise, looking into both what went wrong and the future generation of products that would have been. Unfortunately, these authors were ill informed on the subject, having made errors on the facts and missed key points. This article will attempt to clear up some of the facts. It will not present every single event that occurred at 3dfx, as that would take an entire book. Rather, highlights will be given that took place along the life of the company.

With the initial introduction of the Voodoo Graphics chipset, 3dfx was given a substantial performance lead. As one of the first true 3D accelerators, the competition for it was Rendition’s Verite, S3’s Virge and NVIDIA’s NV1. PowerVR soon followed with a part, but it was plagued with compatibility issues. Even with these competitors, Verite was the only true 3D accelerator, with S3’s decelerator Virge taking a large part of the OEM market. Thus, achieving the performance lead, 3dfx was crowned the winner and the market was theirs.

While it was a product that was not originally scheduled, Voodoo2 soon followed. Voodoo2, much like every other product that followed, was created to fill a gap in 3dfx's product cycles. Voodoo2 again took the market in performance, more than doubling Voodoo Graphic’s performance with SLI configurations. Yet through all this, the goal was to deliver Rampage.



SIDEBAR: The original founders of 3dfx were Gary Tarolli, Scott Sellers, and Ross Smith

Voodoo Banshee

Page:: ( 2 / 6 )

Management changes

It was sometime in between the Voodoo/Voodoo2 period that Greg Ballard came onto the scene as CEO. He was there for marketing, and he was good at it, though there was something missing when it came to technology. He pushed a variety of 3dfx marketing campaigns that helped bring 3dfx to the top. Problems apparently came from his lack of understanding how the graphics industry functioned. Ballard desired to deliver a single chip 2D/3D solution as the competition had thus far done the same.

This would allow 3dfx to enter the mainstream and OEM markets, increasing revenue. It would also renew trust in 3dfx as their ill-fated Voodoo Rush (a multi-chip 2D/3D solution with a separate vendor's 2D core) had created doubters. With limited engineering resources at the time, the only option for this to occur was to remove staff from another project and dedicate them to this. Thus Rampage lost vital engineering resources and Banshee was created.

Voodoo Banshee

With the release of Voodoo Banshee, 3dfx was able to offer a solid 2D/3D solution. Unfortunately, all was not pleasant in the land of 3dfx. Having removed the second texture unit on Banshee's pixel pipeline, multi-texturing performance was below that of a single Voodoo2 solution.

Additionally, NVIDIA for the first time had become a real competitor with their TNT graphics core. NVIDIA's TNT offered similar performance to that of Banshee (in some cases slower, in others faster). Several additional features were built within TNT that were not included in Banshee, such as 32-bit color and textures, as well as high-resolution textures. This made it a somewhat more appealing solution for consumers and developers. 3dfx thus began losing market share and developer confidence.


With all this having taken place and Banshee already having consumed much of Rampage's resources, 3dfx was forced to take Rampage back to the drawing board. It was no longer the high-end board they had hoped. The anticipated market leading performance and feature set no longer existed as NVIDIA had gained considerable ground and the part was simply taking too long to deliver. Thus, with Rampage substantially delayed another stopgap product would be required, this one in the form of Avenger.





SIDEBAR: Greg Ballard is now the Interim CEO at SONICblue (formerly known as S3), one of 3dfx’s original competitors.

Avenger/Delays

Page:: ( 3 / 6 )

Avenger becomes Voodoo3

Avenger, which later became known as Voodoo3, was 3dfx's follow-up to Banshee. Originally this product was to be named Banshee2, for that is really what it was. However, 3dfx management knew that the Voodoo name provided much greater brand recognition and so they opted for that name. Voodoo3’s feature set was identical to that of Banshee. It was simply a higher-clocked version of the previous chip with a second texture unit installed. Performance was definitely competitive, with NVIDIA's TNT2 and TNT2 Ultra often falling behind in performance, but the lack of new features made NVIDIA's solution more appealing once again. This hurt 3dfx's sales and caused them to further lose market share and developers confidence.

Just prior to the launch of Avenger, the merger with STB Systems was announced. STB had been an add-in board manufacturer and they had pretty much dominated the OEM market with products in nearly all the major OEM systems. For 3dfx, the hope was to get their products into OEM systems. For STB the hope was to finally have a say in each chip’s feature set.

Many would say a mistake made by 3dfx in all this was cutting off supply to other board manufactures. With several companies having strong brand recognition in the United States and Europe, this reduced potential sales. Additionally, Asian board makers, typically having niche with Asian system builders, were cut off. This hurt 3dfx’s sales throughout the remainder of their existence.

Product delays

With the oncoming merger almost complete, many at STB were under the impression that 3dfx's next part, Rampage, was all but taped out. This would have been true had 3dfx not decided to make some last minute changes to the design. These were not minor changes either, but major feature introductions. The most important new addition was SLI support. Had SLI not been an included feature, what would be called VSA-100 in its original form, would have been nothing more than a TNT2 Ultra. 3dfx knew this would not be an appealing solution, so Rampage was redesigned to allow for multi-chip boards, theoretically doubling performance (or more, depending on many chips were used). Additionally, 3dfx engineers added FXT-1 texture compression.

Adding technology meant additional delays. Delays not only came from adding features, but also from the new issues that spawned as a result of these additions. Problems crept up along the development path and even more delays were found. Officials within 3dfx did not help this problem either. There were serious delays from simple miscommunications within the company.

One example of this was somebody apparently forgetting to go to Asia to pickup the first batch of completed VSA-100 chips. Another example was a mistake in QA. Quake3 was repeatedly locking their system on Voodoo5 and they could not determine the cause. After a two-week delay the cause was found to be a bad Ghost image that was repeatedly used. These and other reasons set VSA-100 back by weeks.






SIDEBAR: One of the nice perks of working for 3dfx, especially in San Jose, was the

unlimited junk food and drinks. Not only was there a large quantity, but a

huge variety as well. Of course the free lunches were nice too.

GeForce vs. Voodoo5

Page:: ( 4 / 6 )

NVIDIA launches GeForce

While all this was developing, NVIDIA was coming on strong. They had released their GeForce256 chip, which took a nice performance lead over Voodoo3. As a follow up, NVIDIA brought the GeForce2 to market. These two parts offered a considerable number of additional features that 3dfx did not provide with Voodoo5. While 3dfx did offer anti-aliasing that was considerably superior to NVIDIA’s, they had a tough time selling it due to NVIDIA’s aggressive marketing and technology demos. From this, 3dfx lost the majority of their developer support and a considerable amount of consumer confidence.

Voodoo5 6000 problems

In the end, Voodoo5 was a fairly successful product. However, the high-end board, Voodoo5 6000, was forever delayed. There were many happenings with this board, but it boils down to this: 3dfx did not consider the design well enough before the board was announced.

The AGP specification simply was not designed with this type of product in mind. Many attempts were made to work around this, even completely changing the board design and the bridge chip used. Yet in the end, Voodoo5 6000 was canceled in the last weeks of 3dfx.

The specific issue that resulted in the final cancellation was an AGP issue with certain motherboards. While most motherboards did function, there were several that did not quite meet AGP spec, resulting in the boards not functioning. While a BIOS fix on these boards would have likely resolved the issue and though the incompatible boards were few in number, 3dfx chose not release the product. And thus they again failed to retake the performance crown they so badly wanted and lost even more consumer confidence.

While all these events were occurring, 3dfx was losing money. The board manufacturing plant in Mexico was never at capacity, reducing profits on each graphics board sold by roughly 10% from the intended 25% margin. Only in 3dfx’s final months did management decide to start selling out the remaining factory space, filling the product lines. This brought the board plant to near profitability on its own, but this was just one change that was too little, too late.






SIDEBAR: When the announcement was made, most seemed to think the Richardson office

was closing. Few expected the entire company to shut down.

Money/Rampage

Page:: ( 5 / 6 )

More inside details

3dfx was notorious for spending money. In the last year or so, roughly $30-50,000 was spent monthly on lunches. This did not include the additional snacks and drinks that were provided to employees. Hiring didn’t stop until the last few weeks, with all of us keeping hope that the company would pull through. Of course this did not happen.

Could 3dfx have lasted? Perhaps. They were offered a line of credit, but the board opted not to accept it as they would not accept the terms. Rumors within the company also circulated that an investor had expressed strong interest in the company, but backed out from a simple "goof" on the boards’ part (specifically, it was said to be their mention to the investor the possibility of a buyout by another company). But what would the future have held for 3dfx?

Next generation parts

Daytona- 3dfx's first low-end OEM part. Daytona was effectively a VSA-100 part with a DDR memory controller and a 64-bit memory bus. The idea was to deliver a cheaper version of the VSA-100, with the 64-bit bus making a notable dent in cost. Daytona simply could not be finalized though. It would tape out and a bug would be found, then tape out again and another bug would be found. Fortunately, a chip was not made between each tape out with the final number being A7 silicon. In the end, this resulted in considerable delays and final Daytona silicon never coming to life.

[image]

to the fab? Probably not.

Next generation cores

Page:: ( 6 / 6 )

Tantrum- A single chip combination of Rampage and Sage. Targeted at the OEM market, performance would be lower than a Rampage-Sage combination, with considerably reduced cost.

Fear- The first part based on 3dfx and Gigapixel technology. Fear actually consisted of two separate parts: Fusion and Sage II. Fusion was derived from combining 3dfx and Gigapixel technology. This was a part targeted at DirectX8-9 (though the specification was nothing near final). Being from Gigapixel, it was a deferred rendering architecture. At the time of 3dfx closing shop, Fusion was considered RTL complete and tape out was expected in March of 2001. Sage II was slightly behind Fusion, but it was making ground.

Fusion
250+ MHz Core
Approx 60 Million transistors
4 pixel pipelines
8 texture per-pass via loop back
Deferred Rendering Architecture
DX8-DX9 Pixel Shader

Sage2
100 Million Triangles/sec Sustained
300 Million Triangles/sec Theoretical
DX8-DX9 Vertex Shader

Fearless- A single-chip Fusion-Sage2 part. Comparable to what Tantrum was to Rampage.

Mojo- The distant future of 3dfx. This was based on an entirely new generation of design. It was considered the next-generation of deferred rendering. Targeted at DX9 and higher, it had a considerably extensive feature set. With Fear's anticipated performance being such a high level, the raw performance specifications of Mojo were actually slightly lower. Mojo was a single-chip solution unlike Fear and Spectre, including the geometry processor with the pixel pipeline.

Conclusion