FragTek
New member
The Definitive Seven Year GPU History Guide (1995-2002)
Well ol' PV inspired me to go about porting a complete history for the video cards that have paved the path in 3d rendering technology!
We will be going year by year through 3dfx and other makers infamous history as the leader in GPU's until and through the extinction and final run of Voodoo chipsets/cards. (Twas a sad day!
) Let's get right into it!
(All information pulled from Accelenation)
Pre-Voodoo (1995-6)
Some might say that the modern era of 3D game acceleration began with the release of the first Voodoo boards in October 1996 (first demonstrated at E3 in May). It was certainly the case that the gaming world could think of nothing less for the following twelve months. But what of the other 3D cards that were around at this time, and what of the names that have now become synonymous with the graphics industry. Here are a few reminders...
o The Imagine 128 board and chip from Number Nine in 1995.
o The R3D/100 chip from the Lockheed-Martin subsidiary REAL3D in 1995.
o The Diamond Edge 3D board in 1995 sporting NVIDIA's first chip the NV1.
o The Gigi chip from 3Dlabs found on the original Creative 3D Blaster VLB board in 1995.
o The Matrox Mystique board in 1996 utilizing their own MGA-1064SG chip.
o The Hercules Terminator 3D board in 1996 using the ViRGE chip from S3.
o The ATi 3D XPRESSION board in 1996 using their first 3D RAGE chip.
o The second 3D Blaster board from Creative in 1996 based upon Rendition's Vérité1000 chip.
o The Apocalypse 3D from Videologic in 1996 using the PCX chip from NEC/PowerVR.
So you see all the players were there in 1996, and each had a card with at least some degree of 3D functionality. In fact the NV1 from NVIDIA a year earlier was extremely advanced for 1995. Unfortunately NVIDIA had adopted a rendering strategy that was incompatible with the polygon approach incorporated into the new API from Microsoft called DirectX. Support for the product waned and this left the door open for 3Dfx. The following article from ZDNet written in December 1996 managed to capture the mood just after the release of the Voodoo chipset.
The Voodoo boards were unique in their design having been targeted solely for 3D games and incorporated no 2D capability whatsoever. These cards simply switched the output from a second card over to their own whenever a suitable application was run. Many in the industry, including Microsoft, doubted whether people would be prepared to pay the projected 500US$ price tag simply to improve their gaming experience. Whether this belief was founded was never proven because in the spring of 1996 the price of computer memory crashed and the first Voodoo boards eventually retailed at the more agreeable price of 300US$.
To this day there has never been a single impact on PC gaming to equal that delivered by the first Voodoo chip-set. Overnight the gaming world went from a scenery popping, 8-bit, blocky 15fps, to a Z-buffered, 16-bit, texture filtered 30fps. This transformation was demonstrated in no better way than with the very first game to natively support the Voodoo chipset, the newly released Tomb Raider from Core.
At Siggraph in August 1996 Microsoft tried to dictate the future path of 3D rendering technology by announcing their Talisman project. The philosophy behind Talisman was to reduce bandwidth by introducing a tile-based rendering approach similar to the one being developed by PowerVR. Microsoft was supported by no less than Samsung, Fujitsu and Cirrus Logic, all of whom pledged silicon to implement the technology. Unfortunately over the coming months rendering philosophy would be defined by the likes of 3Dfx and not by some vision of the future from Microsoft. One by one each chipmaker pulled out of the project and although some of the technology would eventually be licensed by Trident, the Microsoft dream was dead.
Unfortunately the Talisman project had sidetracked Microsoft and shifted their attention away from the development of DirectX. If Talisman had taken off, history may have witnessed a completely different path for DirectX. As it was, the failure of Talisman meant that it took Microsoft another twelve months to admit defeat, reverse and then forge ahead with the de facto rendering approach. It would not be until the end of 1997 that the world would finally see a competent API from Microsoft in the shape of DirectX version 5. In the absence of a suitable API for games, 3Dfx were left with no choice but to develop their own. Most games from 1996 through 1998 would take advantage of the superior performance resulting from the use of this proprietary API called GLIDE.
At this point I could go on to write a book on the development of graphics over the intervening years, but instead I will try to constrain myself to the more significant events.
Year 1997
One name that isn’t heard much these days in relation to gaming is 3Dlabs. The reason is because they have principally designed boards for the professional workstation market, and these have generally been too expensive for most gamers. A good example was the GLINT 300SX chip announced in April 1994 (shipped that November) and claimed to be the very first 3D hardware accelerated product for the PC. Shortly afterwards 3Dlabs announced their first attempt to bridge the gaming gap with a cut down version called the “Gigi” or “Game Glint” chip. The very first 3D Blaster board from Creative Labs released in August 1995 was based on this chip. The principle crossover from 3Dlabs however came in the form of their Permedia range. The first Permedia chip was announced in October 1995, but it was not until the beginning of 1997 that the first boards appeared. The most memorable of these was the Fire GL 1000 from Diamond which was 30% faster in 3D than a Matrox Millennium II card, but unfortunately still slower and more expensive than a 3Dfx Voodoo card.
3Dfx dominated the market for the first part of 1997 and in February entered into a secret agreement with Sega to produce the chip for their forthcoming Dreamcast console. Unfortunately this would all end in tears when, in September, Sega unexpectedly pulled out of the agreement in favor of using the PowerVR technology from NEC and Videologic. 3Dfx sued for breach of contract and the whole unhappy affair dragged on until the following summer.
ATi shadowed 3Dfx with their Rage II chip for the first quarter of ’97 but in March announced the first real challenge to Voodoo dominance in the form of their improved Rage Pro chip (available in July). Let down only by poor initial drivers (especially for OpenGL) ATi’s Expert@ range of cards were very popular and under DirectX produced both superior performance and almost equal visual quality to that of its rival. Unfortunately for ATi, everyone was just Voodoo-crazy at the time.
In April NVIDIA then entered the fray with their NV3 or RIVA128 chip (available in August) which considerably outperformed the Voodoo under DirectX. This was especially true when using large textures at resolutions beyond 640x480, helped by the RIVA’s wider memory bus and built-in compatibility to DirectX. The RIVA128 even managed to edge out the RAGE PRO from ATi in the performance stakes (although with reduced visual quality), and due to its low cost found favor with system builders.
Also in April Creative Labs shipped their Graphics Blaster 3D board based upon the first 3D chip from Cirrus Logic. The CL-GD5464 or Laguna 3D chip had been announced toward the end of 1996, but in 1997 it faced some stiff competition from 3Dfx, ATi and NVIDIA. The board's low cost failed to compensate for its poor performance, despite its use of RAMBUS memory. Other boards would come to utilize the Laguna chip, but this would prove the last significant 3D venture for the once dominant 2D chipmaker Cirrus Logic.
Trident has produced graphics products since 1987 and is still producing them to this day, mostly for the integration market. Unfortunately Trident is yet another of those lesser companies that have been largely ignored by the media over the years. In May 1997 they shipped their first 3D chip called the 3DImage975. This was a low cost PCI or AGP1X solution with very limited 3D functionality that generally performed no better than a Cirrus Laguna or Trio3D chip from S3. Trident continually tweaked the 975 and toward the end of the year released the 3DImage985 chip with AGP2X and DVD support.
Nothing is certain in this industry. Today NVIDIA look dominant, but then so did a company by the name of Number Nine back in 1995/96. Founded in 1982, narrow profit margins eventually led this board manufacturer to produce its own chips. At Comdex in 1994 the company announced its famous Imagine 128 board that utilized the industry’s first 128bit graphics chip appropriately called “Imagine”. At this time 3D functionality was very basic, but the 128bit memory interface gave the card exceptional 2D performance. This fact, coupled with the card’s high price (999US$), meant that it was targeted solely toward the professional workstation market. In 1996 Number Nine released its Imagine 128 Series 2 board for both PC and Apple MAC platforms. These boards sported the company’s new Imagine 2 chip and retailed for around 700US$. Concerned over the raft of cheaper products released in ’96, Number Nine announced a cut down version of its Series 2 card called the “2e” that retailed for just 399US$. Concurrently Number Nine was also producing boards based on the popular Virge chip from S3 and so in May ’97, when they announced their third generation product, everything looked rosy.
This series 3 chip was called “Ticket to Ride” and was used on their Revolution 3D board that was first demonstrated in June and won the “Best of PC EXPO ‘97” award that same month. In August Number Nine shipped their first AGP version of the board and that year the company was named as one of the fastest growing and most influential in the PC industry. NIVIDA should take note, because today Number Nine no longer exist!
So what went wrong? Was Number Nine beaten into submission by the likes of 3Dfx, ATi and NVIDIA? Well the short answer is yes, but not because of anything that the competition did, rather something that Number Nine did not do; namely texture mapping. Number Nine’s director of communications Phil Parker was asked what would happen if a chip company went one way and Microsoft went another:
"This happened to us. Our Imagine 2 chip didn't have texturing, and the requirement for Direct3D was texturing. We said 'Oh, my god,' but by that time, we couldn't compete with the likes of S3 or NVIDIA."
This lack of texturing was corrected in the T2R chip, but as Phil states, by that time the competition was producing twice the 3D performance and no one seemed to care about 2D image quality any more. Worse still, those professionals that were looking for a high quality yet low cost 2D solution, turned instead to cards like the Permedia 2 that also produced a reasonable 3D performance.
If you were wondering why the chip was called “Ticket to Ride”, this was because most of the products from Number Nine were named in relation to the Beatles. So we got names like Revolution, Imagine, and before that, Pepper, Rita and Tina. There were even secret references hidden within the board BIOS and circuit layout. Another company quirk was its use of controversial ad campaigns as shown above. For these reasons the company was quite popular in its day. Unfortunately its day was done.
Another company that started out in the business sector but increasingly pandered to the gaming community is Matrox. Following the success of their first Millennium board at Comdex in April 1995, the company’s graphics department separated out and became Matrox Graphics Inc. In order to target the gaming community the company announced their Mystique card in June 1996 (available in August) and this was demonstrated at Comdex in November. The MGA-1064SG chip that Matrox used on this board had a slightly reduced specification compared to the original Millennium chip (MGA-2064W). Supporting a maximum of only 4MB of SGRAM as opposed to 8MB of dual-ported video RAM, the card was quickly reduced in price to just 179US$ to make it more attractive to the gaming community.
In May 1997 Matrox announced the Millennium II board for the business market (available in July) and quickly released an AGP version that September. Also in September Matrox announced a board targeted solely to the gaming community called the m3D. This board used the recently announced PCX2 chip from NEC/PowerVR and is the only time that Matrox have used a 3rd party core. This move was especially surprising given that a month earlier, in August, Matrox had already announced its new Mystique 220 gaming card. Matrox probably adopted the PCX2 chip due to the lack of 3D features in its own chips, including the MGA-1164SG found on the new Mystique. At the time, features such as bilinear texture filtering and mip-mapping were included purely for gaming and had therefore not been incorporated into the business oriented products from Matrox. In games both the Millennium II and Mystique 220 boards produced only half the performance of a Voodoo card, with the more capable m3D board coping slightly better.
In June we heard of a new entry in the form of the SiS6326 from Silicon Integrated Systems (SiS). Finally available at the very end of the year it demonstrated a good image quality, but had a dire performance that was nowhere near that of a Voodoo card. However as others would fall by the wayside over the intervening years, SiS would doggedly stick to producing low-end solutions that would eventually be integrated into successful motherboard designs.
One company that had always targeted the 3D gaming market since its formation in January 1993 was Rendition. Its first Vérité 3D chip announced in September ’95 (shipped August ’96) and later renamed as the V1000 was quite impressive for an inaugural product. In August ’97 Rendition announced its V2X00 range of chips. The faster of these, the V2200, first appeared on the Diamond Stealth II card in October and turned in a similar performance to that of a Voodoo card. Unfortunately this was one year after the initial Voodoo release and 3Dfx had not been sitting idle.
3Dfx continued its market dominance throughout 1997 despite some stiff competition from the likes of NVIDIA and ATi. This was due in no small part to its loyal fan base, good visual quality and developer support of its own proprietary API called GLIDE.
In August 3Dfx attempted to incorporate some 2D functionality into their 3D-only product with the announcement of the Voodoo Rush chipset (available in October). I say “3Dfx” attempted, but in reality the 2D part came from Alliance Semiconductor. Effectively this was two cards in one and each seemed to trip over the other. The result was a product with a substandard 2D image quality and a 3D performance that was 10% less than a standard Voodoo board. Although not quite as bad as some would have you believe, the Voodoo Rush was still one product that most 3Dfx fans would rather forget.
Back in May 3Dlabs announced their Permedia 2 chip, and in October one of the first boards to use this was the Fire GL 1000 Pro from Diamond. This board performed on equal terms to a Voodoo board under Direct3D, but as to be expected from a professional card, it excelled under OpenGL. However, at the time, most gamers were content playing games specifically ported to the Voodoo chipset under its own GLIDE API or mini OpenGL driver. The high price of a Permedia 2 card also prevented its widespread uptake, but it did allow many professionals to have a good game of Quake in their lunch break. Some of the initial success of Quake probably derived from its accessibility in the workplace.
In November S3 announced the successor to its popular Virge3D chip. The new 128-bit Trio3D chip would support SGRAM memory running at 125MHz. However, when boards finally shipped in the first half of 1998, their performance fell woefully short of the leading boards for that year. S3 had begun its slide.
In the winter of 1997 3Dfx announced its successor to the Voodoo chip-set, aptly named the Voodoo2 (but not available until the following March). This chip-set included a second texturing unit and a faster clock that effectively doubled its performance and shot it way ahead of the competition. Additionally, the ability to produce quadruple performance by connecting two boards together (SLI) presented a temptation that many testers found difficult to resist. For this reason 3Dfx continued to dominate in league tables throughout 1998 despite the continued presence of some stiff competition.
Year 1998
In February Intel decided to stick its toe into the 3D waters when it announced the i740 chip (available in April). Intel developed the chip in partnership with REAL3D a spin-off from Lockheed-Martin who produced the R3D/100 chip in 1995. In fact Intel had just purchased a 20% minority interest in REAL3D. In games the performance of the i740 was about half that of a single Voodoo2 board, despite 3D Winbench 98 placing it on equal terms. This lackluster performance was a result of Intel using the new AGP bus as it had been intended, for texture rendering. Unfortunately texturing from main memory was much slower than from local memory and this approach never caught on.
Also in February PowerVR/NEC announced their 2nd generation technology for console and PC. However it would not be until December that the PC version would eventually be named as the PVR250 and then a further eight months before the first cards would surface.
NVIDIA were now into their familiar six monthly stride of producing a new product around August and a tweak around February. So in February they announced a tweak on their RIVA128 chip called the RIVA128 ZX (available in May). However, just one month later in March, NVIDIA caused a real stir by announcing the specification of their next wonder-chip the RIVA TNT. The clear intention of NVIDIA was to spoil the Voodoo2 party during 1998, in preparation for the release of this new product in the fall. Unfortunately NVIDIA failed to deliver on the TNT’s projected clock speed and this generated a lot of ill feeling and a reputation for hype. Despite this fact the RIVA TNT still managed to match the performance of a single Voodoo2 board when running under DirectX.
ATi also produced a product tweak in February 1998 in the shape of their 3D RAGE PRO TURBO chip. In reality this was just a re-branding of their previous RAGE PRO chip combined with their latest “TURBO” drivers. ATi's serious Voodoo2 competition came in the form of their RAGE128 GL & VR chips in October of the same year. The faster GL chip managed to match the performance of a single Voodoo2 board, but this was slightly disappointing given the resources of ATi in comparison to either 3Dfx or NVIDIA at the time.
On the 4th May Trident announced that it had shipped over one million 3DImage975 chips. Game players may scoff at products targeted toward the lower end of the market, but that doesn’t mean there isn’t money to be made at that end.
In April Matrox had announced its 64bit G100 chip that was the first Matrox product to include features such as texture filtering and vertex fogging. Then in May the G100 was quickly followed up by the 128bit G200 chip. Of the principle contenders in 1998 the G200 produced the weakest performance, but as expected from Matrox, it had a very solid 2D performance and an excellent all-round image quality.
Back in 1995, or was it 1996, a small Finnish company by the name of Bitboys Oy announced their Pyramid3D chip. As the months rolled by there emerged successive technology reports that detailed such things as a geometry engine (transform and lighting) and projected speeds well in excess of the first Voodoo card from 3Dfx. What was never seen however was the actual product itself. Were BitBoys for real, or was it all a hoax? As I understand it, BitBoys approached a company by the name of VLSI Solutions with a chip design that later turned out to be written on the back of an envelope. TriTech Microelectronics then purchased that envelope with the intention to design the chip themselves. VLSI would then fabricate the board and another company program the drivers. Eventually some pre-production models were produced and demonstrated at Assembly '97, but just before the cards went into mass production in 1998 Tritech pulled out of the project killing it stone dead.
So it would seem that the Pyramid3D project did eventually become legitimate, even if Bitboys never were. The following left-hand image is purported to be a pre-production version of the card, but the right-hand image has been sent in by Mark Vojkovich who actually owns one today. It has 8MB of SDRAM and a pass-though connection similar to that found on a Voodoo board.
In May 1998 BitBoys Oy raised their heads once again and announced their Glaze3D chip. This chip would have a projected performance four times greater than that of the current all-conquering Voodoo2 chipset! At the time this statement caused a considerable stir, especially considering the fiasco surrounding their previous attempt. However once everyone realized that the chip would not be produced for over a year, this interest soon dissipated.
On the 26th May and the 16th June Number Nine announced its forth generation product the Ticket to Ride IV chip and associated Revolution IV graphics card respectively. Despite a conservative price tag of 219US$ for the 32MB model, Number Nine refused to pitch the card to the gaming community, but instead targeted the desktop publishing market where 2D image quality is paramount. To this end the Revolution IV was also bundled with a digital flat panel 1600SW display from SGI. In tests the card did produce a crystal clear image, but unfortunately its 128bit 2D performance was no longer unique and easily matched by cards based on the TNT chip from NVIDIA or the Rage128 chip from ATi. These products also outpaced the Revolution IV card in games, sometimes by more than 200%. Number Nine was pinning its hopes on the 2D market, but this was simply too little too late to save the company.
In July 3Dlabs announced their Permedia 3 chip, but it would not be until May 1999 that the first card would ship. The reason for this delay was probably due to the company's decision to make its own boards — not the only company to make this decision in '98.
Also in July 1998, before either of the two serious offerings from NVIDIA or ATi, it was S3 with their Savage3D chip that was arguably the first to market with a realistic threat to the Voodoo2. Unfortunately the Savage3D drivers took a while to stabilize and S3 sold fewer cards than they had expected, squeezed as they were by the other three principle players. The slide continued.
So what of Rendition in 1998? Well a Vérité V3300 almost saw the light of day but unfortunately the company had financial worries and in June of 1998 it became part of the Micron group. Finally in November Rendition announced that they had shelved the Vérité V3300 and were concentrating instead on a new V4400 chip to be released early in 2000. Yeah promises, promises.
Although 3Dfx seemed unstoppable, the Voodoo products were falling behind in other respects. They required more than one chip. They had no 2D functionality. They did not support the new AGP bus. They were restricted to a 16-bit color output and could handle only small texture sizes. In addition to all of this, they relied heavily upon custom APIs in the form of GLIDE and the mini OpenGL driver that was written specifically for them. The likes of ATi and NVIDIA were already ahead in some of these areas and this fact would become a significant advantage once the Voodoo2 had run its course. The competition was about to pounce.
3Dfx tried to address some of these issues with its Voodoo Banshee released in the fall of 1998. Supposedly based on Voodoo2 technology it did incorporate a 2D core and did plug into an AGP port (although it lacked AGP texturing). Unfortunately the Banshee had one big drawback — it only possessed a single texturing unit. This crippled its performance for the increasing number of games that required dual texturing, most significantly Quake2. By the end of 1998 most of the principle players had a dual texturing capability and so the life span of the Banshee was short to say the least. However despite this fact 3Dfx somehow managed to sell over one million Banshee chips. People just couldn't get enough of 3Dfx!
On 22nd September, having just settled one courtroom battle with Sega, 3Dfx then decided to sue NVIDIA over an alleged patent infringement concerning the implementation of multi-texturing on their TNT chip. As it would turn out this would come to nothing, just as the rabbit that tries to eat your prize carrots, ends up in the pot itself.
On 10th November Trident announced their 128-bit Blade 3D chip. This was the company’s first serious attempt at 3D acceleration and it proved to be a reasonable low-end solution. Although it produced a performance that was 40% less than the TNT chip from NVIDIA, it did produce a good image quality. Unfortunately when cards finally arrived the following March they failed miserably to compete with even the lowest performing solution of ’99. The chip did however match the performance of the i740 from Intel, and was likewise integrated into future motherboard chipsets.
By the end of 1998 there were big changes afoot at 3dfx… and not just a lower-case ‘D’. The company had decided to merge with STB so that it could exclusively manufacture its own boards. At the time people were concerned that this might restrict the distribution and availability of Voodoo products. Regardless, 3dfx marched onward and in November announced its much-anticipated successor to the Voodoo2.
The Voodoo3 specification came as a big disappointment to fans of 3dfx because it offered precious little in the way of additional 3D innovation. It did incorporate a 2D core and produced Voodoo2 SLI performance within a single chip solution. However, as with the Banshee, it still possessed limited AGP compatibility, and came with no support for 32-bit color, no textures greater than 256x256, and no memory configurations higher then 16MB. To compound this, the release of the Voodoo3 would be delayed five months to allow the merger with STB to be finalized. Handing the competition five months on a silver platter was a generous gift indeed.
So 1998 was one hot year for gaming hardware, but it was also one hot year for the games themselves. We began the year with the release of Quake2, which was then followed up by Starcraft, Unreal and Battlezone. Then in the fall came the eagerly awaited Shogo, Sin, Half-Life, Heretic II, Grim Fandango and Thief. Wow what a year! But by the time Christmas had come the market was suffering from heat exhaustion and things turned decidedly chilly. Distributors like EIDOS caught a cold over the Christmas period and 1999 turned into a decidedly poor year for games.
Although 1999 turned out to be a bad year for games, graphics hardware was about to go ballistic!
Year 1999
Amidst much Voodoo3 speculation the year kicked off with the announcement of the SiS300 on the 2nd January. Did anyone notice? The SiS300 was actually a 128bit AGP2X compatible accelerator with support for 64MB of memory, 32bit color, and trilinear filtering. Remember now? Multi-texturing, DX6 compatibility, hardware DVD decoding, TV-out and digital flat-panel support. I must have still been recovering from the New Year’s Eve party, because I don’t remember a thing about it. Demonstrated in April, this chipset (SiS300+SiS301) had excellent visual quality, but due to its single pixel pipeline, had a performance that struggled to match that of even the previous year’s contenders. Despite this fact, SiS would go on to integrate this technology into many of their forthcoming core-logic chips (530,620 & 630). 3D integration would become a trend that many would follow.
On February 16th Number Nine announced the first AGP4X graphics card, the SR9, built around the Savage4 chip from long-term collaborator S3 Graphics (available in May). Also in May they demonstrated the industry’s first DVI interface for the card. However in June Number Nine risked being stranded when S3 announced a shift in its market focus. In August Number Nine announced its intention to pull out of the graphical chip market and concentrate instead on the design of a high-end board built around the Fuzion 150 chip from PixelFusion. The phrase “last gasp” comes to mind. In December Number Nine finally gave up the ghost and sold out to long-term partner S3 Graphics. Five months later in May 2000 it was officially announced that Number Nine had ceased trading. So the company that had given us the first 128bit chip and had been tipped for greatness, did eventually expire.
March saw the announcement of the TNT2, TNT2 Ultra and cheaper Vanta from NVIDIA. In the TNT2 we finally had the chip that NV had promised almost a year earlier. When the cards finally arrived in late April they heralded the end of the dominance of 3dfx. The TNT2 had everything that we have now come to expect from a graphics processor. It had solid 32-bit color and Z-buffer support, good image quality, good drivers, and great performance. The TNT2 Ultra forced even the most ardent of Voodoo devotees to admit that NVIDIA had finally delivered.
In March NEC finally demonstrated its PVR250 chip utilizing the second generation of tile-based rendering technology from PowerVR. First announced some thirteen months earlier and finally named in December ’98, this initial demonstration of the card would merely provide a taster before the NEON 250 from VideoLogic eventually hit the stores in August. In the most part this delay was caused by the phenomenal success of the Sega Dreamcast console, which also happened to use the same chip and consumed them as fast as NEC could bake them. Unfortunately this delay in getting the PC product to market, coupled with a few driver problems, put many people off, despite the PVR250 having near TNT2 performance.
Also in March Matrox announced their G400 board, which turned out to be the top-performing product in '99 until NVIDIA gate-crashed the party in August. The G400 had vastly superior memory bandwidth and produced the best image quality and 32-bit performance on the block. If Matrox had managed to produce more G400MAX versions throughout ’99 the company may have drawn more people away from 3dfx and NVIDIA. Unfortunately, by the time the MAX versions were available, everyone was talking GPU. In the end the G400 sold well, but ironically not to the gaming community, but to the business market where its dual monitor support found favor over its gaming prowess. I felt that the G400 represented a missed opportunity for Matrox, who have since withdrawn back into the business sector from whence they came.
In April ATi announced their RAGE 128 Pro chip (GL & VR), however we had to wait until August for the Rage Fury Pro board that utilized this chip. When it did arrive it managed to match the performance of a standard TNT2 card, but this was too little too late to catch the pre-millennium market. In October ATi announced a board that would equal the performance of even a TNT2 Ultra board. Called the RAGE FURY MAXX, it employed two RAGE 128 Pro chips, but even this level of performance would be insufficient in light of NVIDIA’s autumnal bombshell.
Also in April Intel announced its i752 chip, the successor to its i740 chip released a year earlier. Cards using this chip finally became available in August, but with a performance less than half that of the competition, Intel’s detour outside the world of integration came to an abrupt end.
Back in 1998 S3 produced a very capable chip in the form of their Savage3D, unfortunately the performance of this chip was held back by unstable drivers. You might therefore imagine that the next product from S3 would have these problems sorted and would therefore be a killer chip. On the 1st February this did indeed look likely when S3 announced their Savage4 GT and slightly over-clocked Savage4 Pro chips. Unfortunately by the time boards were available in May it was clear that S3 had missed to boat yet again. Despite producing an excellent image quality the chip seemed to perform no better than ATi’s RAGE 128 or NVIDIA’s TNT and was totally trounced by the TNT2 and Voodoo3 which hit the streets just one month later. In June S3 announced their plans to purchase Diamond Multimedia, probably as a reaction to 3dfx purchasing STB, but without a decent chip, few cared. In August S3 announced their Savage 2000 chip, which did seem quite respectable on paper. Unfortunately there it would remain until the very end of the year.
Back in March the first boards using the Blade 3D chip from Trident appeared. Even though they represented a low-end low-cost solution they still looked pathetic compared to the cards that would be released that year. To correct this, in May Trident announced the Blade 3D Turbo with a boosted clock speed of 150MHz compared to the original 110MHz. This did at least keep the product competitive with the i752 from Intel.
Still in May and almost two years after the first Permedia 2 board hit the streets, 3Dlabs finally shipped their Permedia 3 Create! board. In games the new board performed about 10% less than the TNT chip from NVIDIA or the Rage128 chip from ATi. However in professional packages such as 3D Studio Max where the rendering time of more complex single frames is important, the workstation heritage of 3Dlabs shone through. In such applications no other card released in 1999 could beat it on performance under OpenGL or Direct3D. This disparity between gaming and professional performance clearly highlights the difference between each discipline. In games the focus is on maximizing the performance of comparatively simple images that use a limited number of functions under Direct3D or OpenGL. However professionals render more complex scenes and therefore require compatibility to a much broader range of rendering functions. It is this robust driver, software and user support that makes professional products more expensive compared to their gaming counterparts. Where the same manufacture is concerned the hardware can actually be identical between the two markets. As for the Permedia 3 it basically fell between the two stools and was best suited to professionals who liked to play the odd computer game. To date, this has been the last foray into the gaming sector by 3Dlabs, however they are still forging ahead successfully in the workstation market.
At Siggraph99 in August the Bitboys were back. This time with an updated Glaze3D specification that included every feature under the sun, including 9MB of embedded DRAM memory and four pixel pipelines capable of rendering 600Mpixels/sec and 1.2Gtexels/sec. In addition to this, the new chips could be connected in parallel to produce a phenomenal 1.2Gpixel/sec and 2.4Gtexel/sec. Was it merely a coincidence that this specification seemed to mirror many of the forthcoming features from NVIDIA and 3dfx. Would we ever see a Glaze3D chip? Don't hold your breath .
Okay, so I have now cleared the ground for arguably the two most significant releases in 1999. First the Voodoo3 from 3dfx.
Today most producers try to release a product shortly after it is announced. However a few years ago this was not the case and often a considerable period of time would elapse before an announced product would actually hit the streets. The Voodoo3 was a case in point. It was first announced in November 1998 in order to keep to the traditional date for a Voodoo release. The company even demonstrated the new product at Comdex98, however due to the merger with STB and in order to avoid the premature disruption of Banshee sales, the Voodoo3 was withheld until April 1999.
3dfx was one of the first manufactures to re-label their product range in order to target different sectors of the market. So instead of getting a Voodoo3 we were treated to a Voodoo2000, 3000 & 3500 (sound familiar?). Sadly the five-month delay in their release had allowed the competition to catch up and for this reason the new range failed to inspire. It performed well enough, however this performance derived from a legacy of direct industry support via the GLIDE and mini-OpenGL interfaces (APIs). Unfortunately 1999 would witness both the demise of GLIDE and the introduction of competent OpenGL support from the rest of the industry. Strip these away and suddenly the Voodoo3 range was no better than competing cards such as the NVIDIA TNT2 or Matrox G400 released shortly after. In fact the lack of 32-bit color support and AGP texturing led some to conclude that it was 3dfx that needed to catch up. Despite these shortcomings the Voodoo3 sold well, but not quite as well as the TNT2 from NVIDIA.
This unfamiliar situation that 3dfx found themselves in would be further exacerbated by the second significant release in 1999; namely the GeForce256 from NVIDIA. First hinted at in April, but announced in August and demonstrated at ECTS in September, the GeForce256 captured the imagination of the industry by performing the transformation and lighting calculations in hardware (know as Hardware T&L or TnL). Thus the humble graphics chip had been elevated to the status of GPU or Graphical Processing Unit, and although it would be a couple of months before the first retail cards would surface, this innovation alone managed to steal most of the limelight.
3dfx tried to counter this by simultaneously announcing their new parallel T-buffer technology intended for their forthcoming Voodoo4 range. But lack of actual silicon betrayed the fact that 3dfx were nowhere near ready for a release and NVIDIA stole the show.
Ironically, although it was the TnL engine that caught the public imagination, in the end it was the superior fill rate of the GeForce256 that finally established the chip. NVIDIA already had the best selling and one of the best performing chips on the market in the form of their TNT2 chip (equaled by the Voodoo3 but beaten by the G400). The GeForce256 was like two of these chips strapped together. Its ability to render four single-textured pixels in one pass resulted in some impressive scores, sometimes exceeding the performance of a Voodoo3 or G400 by as much as 40%. In December, when DDR memory arrived, this performance gap was further increased to 75%. These performance figures alone held the attention of the market until the following April, when NVIDIA unleashed their second killer chip.
As previously mentioned, August saw S3 announce their Savage2000 chip. In November the first card using this chip was released and came in the form of the Diamond Viper II. The story of the Savage2000 from S3 demonstrates quite clearly the impact of NVIDIA. Everyone had been waiting for the Savage2000 and when it finally arrived it had the second best performance on the street (about equal to a Matrox G400MAX, but slightly less than a GeForce256). However over the following few months the Savage2000 would fade into obscurity. Although troubled TnL support was partly responsible for this, the principle reason was simply bad timing. At any other time the Savage2000 would have succeeded, but at this time everyone was talking about GeForce, Hardware TnL, and the future of 3dfx. Few noticed the Savage2000, and those that did soon forgot about it.
In October Lockheed-Martin closed down REAL3D and sold all remaining assets to Intel who already owned a 20% stake.
At Comdex in November 3dfx finally named their forthcoming products. The cards would be called the Voodoo4 4500, Voodoo5 5000, Voodoo5 5500 & Voodoo5 6000. Each product would utilize one, two or four of their VSA100 chips (Voodoo Scalable Architecture). It was this parallel design that would implement the T-Buffer technology that had been previously announced at ECTS. The principle feature of this new parallel approach was the promise of full and backwards compatible antialiasing at realistic frame rates. Unfortunately this would require more memory than was commonly used and to make the situation worse memory prices had started to climb. This was an ironic twist, because it was the initial fall of memory prices in 1996 that had contributed to the company’s initial success. The development of a modular approach also led to instability problems. This all resulted in the final release being delayed until the following June and then only after a recall. Having to wait five months before releasing their Voodoo3000 range was bad enough, but now having to sit back a further seven months and watch NVIDIA climbed to the top of the hill, must have been totally demoralizing. Could things possibly get any worse for the once golden boys of gaming? You bet your sweat bippy they could!
Well ol' PV inspired me to go about porting a complete history for the video cards that have paved the path in 3d rendering technology!
We will be going year by year through 3dfx and other makers infamous history as the leader in GPU's until and through the extinction and final run of Voodoo chipsets/cards. (Twas a sad day!
) Let's get right into it!(All information pulled from Accelenation)
Pre-Voodoo (1995-6)
Some might say that the modern era of 3D game acceleration began with the release of the first Voodoo boards in October 1996 (first demonstrated at E3 in May). It was certainly the case that the gaming world could think of nothing less for the following twelve months. But what of the other 3D cards that were around at this time, and what of the names that have now become synonymous with the graphics industry. Here are a few reminders...
o The Imagine 128 board and chip from Number Nine in 1995.
o The R3D/100 chip from the Lockheed-Martin subsidiary REAL3D in 1995.
o The Diamond Edge 3D board in 1995 sporting NVIDIA's first chip the NV1.
o The Gigi chip from 3Dlabs found on the original Creative 3D Blaster VLB board in 1995.
o The Matrox Mystique board in 1996 utilizing their own MGA-1064SG chip.
o The Hercules Terminator 3D board in 1996 using the ViRGE chip from S3.
o The ATi 3D XPRESSION board in 1996 using their first 3D RAGE chip.
o The second 3D Blaster board from Creative in 1996 based upon Rendition's Vérité1000 chip.
o The Apocalypse 3D from Videologic in 1996 using the PCX chip from NEC/PowerVR.
So you see all the players were there in 1996, and each had a card with at least some degree of 3D functionality. In fact the NV1 from NVIDIA a year earlier was extremely advanced for 1995. Unfortunately NVIDIA had adopted a rendering strategy that was incompatible with the polygon approach incorporated into the new API from Microsoft called DirectX. Support for the product waned and this left the door open for 3Dfx. The following article from ZDNet written in December 1996 managed to capture the mood just after the release of the Voodoo chipset.
The Voodoo boards were unique in their design having been targeted solely for 3D games and incorporated no 2D capability whatsoever. These cards simply switched the output from a second card over to their own whenever a suitable application was run. Many in the industry, including Microsoft, doubted whether people would be prepared to pay the projected 500US$ price tag simply to improve their gaming experience. Whether this belief was founded was never proven because in the spring of 1996 the price of computer memory crashed and the first Voodoo boards eventually retailed at the more agreeable price of 300US$.
To this day there has never been a single impact on PC gaming to equal that delivered by the first Voodoo chip-set. Overnight the gaming world went from a scenery popping, 8-bit, blocky 15fps, to a Z-buffered, 16-bit, texture filtered 30fps. This transformation was demonstrated in no better way than with the very first game to natively support the Voodoo chipset, the newly released Tomb Raider from Core.
At Siggraph in August 1996 Microsoft tried to dictate the future path of 3D rendering technology by announcing their Talisman project. The philosophy behind Talisman was to reduce bandwidth by introducing a tile-based rendering approach similar to the one being developed by PowerVR. Microsoft was supported by no less than Samsung, Fujitsu and Cirrus Logic, all of whom pledged silicon to implement the technology. Unfortunately over the coming months rendering philosophy would be defined by the likes of 3Dfx and not by some vision of the future from Microsoft. One by one each chipmaker pulled out of the project and although some of the technology would eventually be licensed by Trident, the Microsoft dream was dead.
Unfortunately the Talisman project had sidetracked Microsoft and shifted their attention away from the development of DirectX. If Talisman had taken off, history may have witnessed a completely different path for DirectX. As it was, the failure of Talisman meant that it took Microsoft another twelve months to admit defeat, reverse and then forge ahead with the de facto rendering approach. It would not be until the end of 1997 that the world would finally see a competent API from Microsoft in the shape of DirectX version 5. In the absence of a suitable API for games, 3Dfx were left with no choice but to develop their own. Most games from 1996 through 1998 would take advantage of the superior performance resulting from the use of this proprietary API called GLIDE.
At this point I could go on to write a book on the development of graphics over the intervening years, but instead I will try to constrain myself to the more significant events.
Year 1997
One name that isn’t heard much these days in relation to gaming is 3Dlabs. The reason is because they have principally designed boards for the professional workstation market, and these have generally been too expensive for most gamers. A good example was the GLINT 300SX chip announced in April 1994 (shipped that November) and claimed to be the very first 3D hardware accelerated product for the PC. Shortly afterwards 3Dlabs announced their first attempt to bridge the gaming gap with a cut down version called the “Gigi” or “Game Glint” chip. The very first 3D Blaster board from Creative Labs released in August 1995 was based on this chip. The principle crossover from 3Dlabs however came in the form of their Permedia range. The first Permedia chip was announced in October 1995, but it was not until the beginning of 1997 that the first boards appeared. The most memorable of these was the Fire GL 1000 from Diamond which was 30% faster in 3D than a Matrox Millennium II card, but unfortunately still slower and more expensive than a 3Dfx Voodoo card.
3Dfx dominated the market for the first part of 1997 and in February entered into a secret agreement with Sega to produce the chip for their forthcoming Dreamcast console. Unfortunately this would all end in tears when, in September, Sega unexpectedly pulled out of the agreement in favor of using the PowerVR technology from NEC and Videologic. 3Dfx sued for breach of contract and the whole unhappy affair dragged on until the following summer.
ATi shadowed 3Dfx with their Rage II chip for the first quarter of ’97 but in March announced the first real challenge to Voodoo dominance in the form of their improved Rage Pro chip (available in July). Let down only by poor initial drivers (especially for OpenGL) ATi’s Expert@ range of cards were very popular and under DirectX produced both superior performance and almost equal visual quality to that of its rival. Unfortunately for ATi, everyone was just Voodoo-crazy at the time.
In April NVIDIA then entered the fray with their NV3 or RIVA128 chip (available in August) which considerably outperformed the Voodoo under DirectX. This was especially true when using large textures at resolutions beyond 640x480, helped by the RIVA’s wider memory bus and built-in compatibility to DirectX. The RIVA128 even managed to edge out the RAGE PRO from ATi in the performance stakes (although with reduced visual quality), and due to its low cost found favor with system builders.
Also in April Creative Labs shipped their Graphics Blaster 3D board based upon the first 3D chip from Cirrus Logic. The CL-GD5464 or Laguna 3D chip had been announced toward the end of 1996, but in 1997 it faced some stiff competition from 3Dfx, ATi and NVIDIA. The board's low cost failed to compensate for its poor performance, despite its use of RAMBUS memory. Other boards would come to utilize the Laguna chip, but this would prove the last significant 3D venture for the once dominant 2D chipmaker Cirrus Logic.
Trident has produced graphics products since 1987 and is still producing them to this day, mostly for the integration market. Unfortunately Trident is yet another of those lesser companies that have been largely ignored by the media over the years. In May 1997 they shipped their first 3D chip called the 3DImage975. This was a low cost PCI or AGP1X solution with very limited 3D functionality that generally performed no better than a Cirrus Laguna or Trio3D chip from S3. Trident continually tweaked the 975 and toward the end of the year released the 3DImage985 chip with AGP2X and DVD support.
Nothing is certain in this industry. Today NVIDIA look dominant, but then so did a company by the name of Number Nine back in 1995/96. Founded in 1982, narrow profit margins eventually led this board manufacturer to produce its own chips. At Comdex in 1994 the company announced its famous Imagine 128 board that utilized the industry’s first 128bit graphics chip appropriately called “Imagine”. At this time 3D functionality was very basic, but the 128bit memory interface gave the card exceptional 2D performance. This fact, coupled with the card’s high price (999US$), meant that it was targeted solely toward the professional workstation market. In 1996 Number Nine released its Imagine 128 Series 2 board for both PC and Apple MAC platforms. These boards sported the company’s new Imagine 2 chip and retailed for around 700US$. Concerned over the raft of cheaper products released in ’96, Number Nine announced a cut down version of its Series 2 card called the “2e” that retailed for just 399US$. Concurrently Number Nine was also producing boards based on the popular Virge chip from S3 and so in May ’97, when they announced their third generation product, everything looked rosy.
This series 3 chip was called “Ticket to Ride” and was used on their Revolution 3D board that was first demonstrated in June and won the “Best of PC EXPO ‘97” award that same month. In August Number Nine shipped their first AGP version of the board and that year the company was named as one of the fastest growing and most influential in the PC industry. NIVIDA should take note, because today Number Nine no longer exist!
So what went wrong? Was Number Nine beaten into submission by the likes of 3Dfx, ATi and NVIDIA? Well the short answer is yes, but not because of anything that the competition did, rather something that Number Nine did not do; namely texture mapping. Number Nine’s director of communications Phil Parker was asked what would happen if a chip company went one way and Microsoft went another:
"This happened to us. Our Imagine 2 chip didn't have texturing, and the requirement for Direct3D was texturing. We said 'Oh, my god,' but by that time, we couldn't compete with the likes of S3 or NVIDIA."
This lack of texturing was corrected in the T2R chip, but as Phil states, by that time the competition was producing twice the 3D performance and no one seemed to care about 2D image quality any more. Worse still, those professionals that were looking for a high quality yet low cost 2D solution, turned instead to cards like the Permedia 2 that also produced a reasonable 3D performance.
If you were wondering why the chip was called “Ticket to Ride”, this was because most of the products from Number Nine were named in relation to the Beatles. So we got names like Revolution, Imagine, and before that, Pepper, Rita and Tina. There were even secret references hidden within the board BIOS and circuit layout. Another company quirk was its use of controversial ad campaigns as shown above. For these reasons the company was quite popular in its day. Unfortunately its day was done.
Another company that started out in the business sector but increasingly pandered to the gaming community is Matrox. Following the success of their first Millennium board at Comdex in April 1995, the company’s graphics department separated out and became Matrox Graphics Inc. In order to target the gaming community the company announced their Mystique card in June 1996 (available in August) and this was demonstrated at Comdex in November. The MGA-1064SG chip that Matrox used on this board had a slightly reduced specification compared to the original Millennium chip (MGA-2064W). Supporting a maximum of only 4MB of SGRAM as opposed to 8MB of dual-ported video RAM, the card was quickly reduced in price to just 179US$ to make it more attractive to the gaming community.
In May 1997 Matrox announced the Millennium II board for the business market (available in July) and quickly released an AGP version that September. Also in September Matrox announced a board targeted solely to the gaming community called the m3D. This board used the recently announced PCX2 chip from NEC/PowerVR and is the only time that Matrox have used a 3rd party core. This move was especially surprising given that a month earlier, in August, Matrox had already announced its new Mystique 220 gaming card. Matrox probably adopted the PCX2 chip due to the lack of 3D features in its own chips, including the MGA-1164SG found on the new Mystique. At the time, features such as bilinear texture filtering and mip-mapping were included purely for gaming and had therefore not been incorporated into the business oriented products from Matrox. In games both the Millennium II and Mystique 220 boards produced only half the performance of a Voodoo card, with the more capable m3D board coping slightly better.
In June we heard of a new entry in the form of the SiS6326 from Silicon Integrated Systems (SiS). Finally available at the very end of the year it demonstrated a good image quality, but had a dire performance that was nowhere near that of a Voodoo card. However as others would fall by the wayside over the intervening years, SiS would doggedly stick to producing low-end solutions that would eventually be integrated into successful motherboard designs.
One company that had always targeted the 3D gaming market since its formation in January 1993 was Rendition. Its first Vérité 3D chip announced in September ’95 (shipped August ’96) and later renamed as the V1000 was quite impressive for an inaugural product. In August ’97 Rendition announced its V2X00 range of chips. The faster of these, the V2200, first appeared on the Diamond Stealth II card in October and turned in a similar performance to that of a Voodoo card. Unfortunately this was one year after the initial Voodoo release and 3Dfx had not been sitting idle.
3Dfx continued its market dominance throughout 1997 despite some stiff competition from the likes of NVIDIA and ATi. This was due in no small part to its loyal fan base, good visual quality and developer support of its own proprietary API called GLIDE.
In August 3Dfx attempted to incorporate some 2D functionality into their 3D-only product with the announcement of the Voodoo Rush chipset (available in October). I say “3Dfx” attempted, but in reality the 2D part came from Alliance Semiconductor. Effectively this was two cards in one and each seemed to trip over the other. The result was a product with a substandard 2D image quality and a 3D performance that was 10% less than a standard Voodoo board. Although not quite as bad as some would have you believe, the Voodoo Rush was still one product that most 3Dfx fans would rather forget.
Back in May 3Dlabs announced their Permedia 2 chip, and in October one of the first boards to use this was the Fire GL 1000 Pro from Diamond. This board performed on equal terms to a Voodoo board under Direct3D, but as to be expected from a professional card, it excelled under OpenGL. However, at the time, most gamers were content playing games specifically ported to the Voodoo chipset under its own GLIDE API or mini OpenGL driver. The high price of a Permedia 2 card also prevented its widespread uptake, but it did allow many professionals to have a good game of Quake in their lunch break. Some of the initial success of Quake probably derived from its accessibility in the workplace.
In November S3 announced the successor to its popular Virge3D chip. The new 128-bit Trio3D chip would support SGRAM memory running at 125MHz. However, when boards finally shipped in the first half of 1998, their performance fell woefully short of the leading boards for that year. S3 had begun its slide.
In the winter of 1997 3Dfx announced its successor to the Voodoo chip-set, aptly named the Voodoo2 (but not available until the following March). This chip-set included a second texturing unit and a faster clock that effectively doubled its performance and shot it way ahead of the competition. Additionally, the ability to produce quadruple performance by connecting two boards together (SLI) presented a temptation that many testers found difficult to resist. For this reason 3Dfx continued to dominate in league tables throughout 1998 despite the continued presence of some stiff competition.
Year 1998
In February Intel decided to stick its toe into the 3D waters when it announced the i740 chip (available in April). Intel developed the chip in partnership with REAL3D a spin-off from Lockheed-Martin who produced the R3D/100 chip in 1995. In fact Intel had just purchased a 20% minority interest in REAL3D. In games the performance of the i740 was about half that of a single Voodoo2 board, despite 3D Winbench 98 placing it on equal terms. This lackluster performance was a result of Intel using the new AGP bus as it had been intended, for texture rendering. Unfortunately texturing from main memory was much slower than from local memory and this approach never caught on.
Also in February PowerVR/NEC announced their 2nd generation technology for console and PC. However it would not be until December that the PC version would eventually be named as the PVR250 and then a further eight months before the first cards would surface.
NVIDIA were now into their familiar six monthly stride of producing a new product around August and a tweak around February. So in February they announced a tweak on their RIVA128 chip called the RIVA128 ZX (available in May). However, just one month later in March, NVIDIA caused a real stir by announcing the specification of their next wonder-chip the RIVA TNT. The clear intention of NVIDIA was to spoil the Voodoo2 party during 1998, in preparation for the release of this new product in the fall. Unfortunately NVIDIA failed to deliver on the TNT’s projected clock speed and this generated a lot of ill feeling and a reputation for hype. Despite this fact the RIVA TNT still managed to match the performance of a single Voodoo2 board when running under DirectX.
ATi also produced a product tweak in February 1998 in the shape of their 3D RAGE PRO TURBO chip. In reality this was just a re-branding of their previous RAGE PRO chip combined with their latest “TURBO” drivers. ATi's serious Voodoo2 competition came in the form of their RAGE128 GL & VR chips in October of the same year. The faster GL chip managed to match the performance of a single Voodoo2 board, but this was slightly disappointing given the resources of ATi in comparison to either 3Dfx or NVIDIA at the time.
On the 4th May Trident announced that it had shipped over one million 3DImage975 chips. Game players may scoff at products targeted toward the lower end of the market, but that doesn’t mean there isn’t money to be made at that end.
In April Matrox had announced its 64bit G100 chip that was the first Matrox product to include features such as texture filtering and vertex fogging. Then in May the G100 was quickly followed up by the 128bit G200 chip. Of the principle contenders in 1998 the G200 produced the weakest performance, but as expected from Matrox, it had a very solid 2D performance and an excellent all-round image quality.
Back in 1995, or was it 1996, a small Finnish company by the name of Bitboys Oy announced their Pyramid3D chip. As the months rolled by there emerged successive technology reports that detailed such things as a geometry engine (transform and lighting) and projected speeds well in excess of the first Voodoo card from 3Dfx. What was never seen however was the actual product itself. Were BitBoys for real, or was it all a hoax? As I understand it, BitBoys approached a company by the name of VLSI Solutions with a chip design that later turned out to be written on the back of an envelope. TriTech Microelectronics then purchased that envelope with the intention to design the chip themselves. VLSI would then fabricate the board and another company program the drivers. Eventually some pre-production models were produced and demonstrated at Assembly '97, but just before the cards went into mass production in 1998 Tritech pulled out of the project killing it stone dead.
So it would seem that the Pyramid3D project did eventually become legitimate, even if Bitboys never were. The following left-hand image is purported to be a pre-production version of the card, but the right-hand image has been sent in by Mark Vojkovich who actually owns one today. It has 8MB of SDRAM and a pass-though connection similar to that found on a Voodoo board.
In May 1998 BitBoys Oy raised their heads once again and announced their Glaze3D chip. This chip would have a projected performance four times greater than that of the current all-conquering Voodoo2 chipset! At the time this statement caused a considerable stir, especially considering the fiasco surrounding their previous attempt. However once everyone realized that the chip would not be produced for over a year, this interest soon dissipated.
On the 26th May and the 16th June Number Nine announced its forth generation product the Ticket to Ride IV chip and associated Revolution IV graphics card respectively. Despite a conservative price tag of 219US$ for the 32MB model, Number Nine refused to pitch the card to the gaming community, but instead targeted the desktop publishing market where 2D image quality is paramount. To this end the Revolution IV was also bundled with a digital flat panel 1600SW display from SGI. In tests the card did produce a crystal clear image, but unfortunately its 128bit 2D performance was no longer unique and easily matched by cards based on the TNT chip from NVIDIA or the Rage128 chip from ATi. These products also outpaced the Revolution IV card in games, sometimes by more than 200%. Number Nine was pinning its hopes on the 2D market, but this was simply too little too late to save the company.
In July 3Dlabs announced their Permedia 3 chip, but it would not be until May 1999 that the first card would ship. The reason for this delay was probably due to the company's decision to make its own boards — not the only company to make this decision in '98.
Also in July 1998, before either of the two serious offerings from NVIDIA or ATi, it was S3 with their Savage3D chip that was arguably the first to market with a realistic threat to the Voodoo2. Unfortunately the Savage3D drivers took a while to stabilize and S3 sold fewer cards than they had expected, squeezed as they were by the other three principle players. The slide continued.
So what of Rendition in 1998? Well a Vérité V3300 almost saw the light of day but unfortunately the company had financial worries and in June of 1998 it became part of the Micron group. Finally in November Rendition announced that they had shelved the Vérité V3300 and were concentrating instead on a new V4400 chip to be released early in 2000. Yeah promises, promises.
Although 3Dfx seemed unstoppable, the Voodoo products were falling behind in other respects. They required more than one chip. They had no 2D functionality. They did not support the new AGP bus. They were restricted to a 16-bit color output and could handle only small texture sizes. In addition to all of this, they relied heavily upon custom APIs in the form of GLIDE and the mini OpenGL driver that was written specifically for them. The likes of ATi and NVIDIA were already ahead in some of these areas and this fact would become a significant advantage once the Voodoo2 had run its course. The competition was about to pounce.
3Dfx tried to address some of these issues with its Voodoo Banshee released in the fall of 1998. Supposedly based on Voodoo2 technology it did incorporate a 2D core and did plug into an AGP port (although it lacked AGP texturing). Unfortunately the Banshee had one big drawback — it only possessed a single texturing unit. This crippled its performance for the increasing number of games that required dual texturing, most significantly Quake2. By the end of 1998 most of the principle players had a dual texturing capability and so the life span of the Banshee was short to say the least. However despite this fact 3Dfx somehow managed to sell over one million Banshee chips. People just couldn't get enough of 3Dfx!
On 22nd September, having just settled one courtroom battle with Sega, 3Dfx then decided to sue NVIDIA over an alleged patent infringement concerning the implementation of multi-texturing on their TNT chip. As it would turn out this would come to nothing, just as the rabbit that tries to eat your prize carrots, ends up in the pot itself.
On 10th November Trident announced their 128-bit Blade 3D chip. This was the company’s first serious attempt at 3D acceleration and it proved to be a reasonable low-end solution. Although it produced a performance that was 40% less than the TNT chip from NVIDIA, it did produce a good image quality. Unfortunately when cards finally arrived the following March they failed miserably to compete with even the lowest performing solution of ’99. The chip did however match the performance of the i740 from Intel, and was likewise integrated into future motherboard chipsets.
By the end of 1998 there were big changes afoot at 3dfx… and not just a lower-case ‘D’. The company had decided to merge with STB so that it could exclusively manufacture its own boards. At the time people were concerned that this might restrict the distribution and availability of Voodoo products. Regardless, 3dfx marched onward and in November announced its much-anticipated successor to the Voodoo2.
The Voodoo3 specification came as a big disappointment to fans of 3dfx because it offered precious little in the way of additional 3D innovation. It did incorporate a 2D core and produced Voodoo2 SLI performance within a single chip solution. However, as with the Banshee, it still possessed limited AGP compatibility, and came with no support for 32-bit color, no textures greater than 256x256, and no memory configurations higher then 16MB. To compound this, the release of the Voodoo3 would be delayed five months to allow the merger with STB to be finalized. Handing the competition five months on a silver platter was a generous gift indeed.
So 1998 was one hot year for gaming hardware, but it was also one hot year for the games themselves. We began the year with the release of Quake2, which was then followed up by Starcraft, Unreal and Battlezone. Then in the fall came the eagerly awaited Shogo, Sin, Half-Life, Heretic II, Grim Fandango and Thief. Wow what a year! But by the time Christmas had come the market was suffering from heat exhaustion and things turned decidedly chilly. Distributors like EIDOS caught a cold over the Christmas period and 1999 turned into a decidedly poor year for games.
Although 1999 turned out to be a bad year for games, graphics hardware was about to go ballistic!
Year 1999
Amidst much Voodoo3 speculation the year kicked off with the announcement of the SiS300 on the 2nd January. Did anyone notice? The SiS300 was actually a 128bit AGP2X compatible accelerator with support for 64MB of memory, 32bit color, and trilinear filtering. Remember now? Multi-texturing, DX6 compatibility, hardware DVD decoding, TV-out and digital flat-panel support. I must have still been recovering from the New Year’s Eve party, because I don’t remember a thing about it. Demonstrated in April, this chipset (SiS300+SiS301) had excellent visual quality, but due to its single pixel pipeline, had a performance that struggled to match that of even the previous year’s contenders. Despite this fact, SiS would go on to integrate this technology into many of their forthcoming core-logic chips (530,620 & 630). 3D integration would become a trend that many would follow.
On February 16th Number Nine announced the first AGP4X graphics card, the SR9, built around the Savage4 chip from long-term collaborator S3 Graphics (available in May). Also in May they demonstrated the industry’s first DVI interface for the card. However in June Number Nine risked being stranded when S3 announced a shift in its market focus. In August Number Nine announced its intention to pull out of the graphical chip market and concentrate instead on the design of a high-end board built around the Fuzion 150 chip from PixelFusion. The phrase “last gasp” comes to mind. In December Number Nine finally gave up the ghost and sold out to long-term partner S3 Graphics. Five months later in May 2000 it was officially announced that Number Nine had ceased trading. So the company that had given us the first 128bit chip and had been tipped for greatness, did eventually expire.
March saw the announcement of the TNT2, TNT2 Ultra and cheaper Vanta from NVIDIA. In the TNT2 we finally had the chip that NV had promised almost a year earlier. When the cards finally arrived in late April they heralded the end of the dominance of 3dfx. The TNT2 had everything that we have now come to expect from a graphics processor. It had solid 32-bit color and Z-buffer support, good image quality, good drivers, and great performance. The TNT2 Ultra forced even the most ardent of Voodoo devotees to admit that NVIDIA had finally delivered.
In March NEC finally demonstrated its PVR250 chip utilizing the second generation of tile-based rendering technology from PowerVR. First announced some thirteen months earlier and finally named in December ’98, this initial demonstration of the card would merely provide a taster before the NEON 250 from VideoLogic eventually hit the stores in August. In the most part this delay was caused by the phenomenal success of the Sega Dreamcast console, which also happened to use the same chip and consumed them as fast as NEC could bake them. Unfortunately this delay in getting the PC product to market, coupled with a few driver problems, put many people off, despite the PVR250 having near TNT2 performance.
Also in March Matrox announced their G400 board, which turned out to be the top-performing product in '99 until NVIDIA gate-crashed the party in August. The G400 had vastly superior memory bandwidth and produced the best image quality and 32-bit performance on the block. If Matrox had managed to produce more G400MAX versions throughout ’99 the company may have drawn more people away from 3dfx and NVIDIA. Unfortunately, by the time the MAX versions were available, everyone was talking GPU. In the end the G400 sold well, but ironically not to the gaming community, but to the business market where its dual monitor support found favor over its gaming prowess. I felt that the G400 represented a missed opportunity for Matrox, who have since withdrawn back into the business sector from whence they came.
In April ATi announced their RAGE 128 Pro chip (GL & VR), however we had to wait until August for the Rage Fury Pro board that utilized this chip. When it did arrive it managed to match the performance of a standard TNT2 card, but this was too little too late to catch the pre-millennium market. In October ATi announced a board that would equal the performance of even a TNT2 Ultra board. Called the RAGE FURY MAXX, it employed two RAGE 128 Pro chips, but even this level of performance would be insufficient in light of NVIDIA’s autumnal bombshell.
Also in April Intel announced its i752 chip, the successor to its i740 chip released a year earlier. Cards using this chip finally became available in August, but with a performance less than half that of the competition, Intel’s detour outside the world of integration came to an abrupt end.
Back in 1998 S3 produced a very capable chip in the form of their Savage3D, unfortunately the performance of this chip was held back by unstable drivers. You might therefore imagine that the next product from S3 would have these problems sorted and would therefore be a killer chip. On the 1st February this did indeed look likely when S3 announced their Savage4 GT and slightly over-clocked Savage4 Pro chips. Unfortunately by the time boards were available in May it was clear that S3 had missed to boat yet again. Despite producing an excellent image quality the chip seemed to perform no better than ATi’s RAGE 128 or NVIDIA’s TNT and was totally trounced by the TNT2 and Voodoo3 which hit the streets just one month later. In June S3 announced their plans to purchase Diamond Multimedia, probably as a reaction to 3dfx purchasing STB, but without a decent chip, few cared. In August S3 announced their Savage 2000 chip, which did seem quite respectable on paper. Unfortunately there it would remain until the very end of the year.
Back in March the first boards using the Blade 3D chip from Trident appeared. Even though they represented a low-end low-cost solution they still looked pathetic compared to the cards that would be released that year. To correct this, in May Trident announced the Blade 3D Turbo with a boosted clock speed of 150MHz compared to the original 110MHz. This did at least keep the product competitive with the i752 from Intel.
Still in May and almost two years after the first Permedia 2 board hit the streets, 3Dlabs finally shipped their Permedia 3 Create! board. In games the new board performed about 10% less than the TNT chip from NVIDIA or the Rage128 chip from ATi. However in professional packages such as 3D Studio Max where the rendering time of more complex single frames is important, the workstation heritage of 3Dlabs shone through. In such applications no other card released in 1999 could beat it on performance under OpenGL or Direct3D. This disparity between gaming and professional performance clearly highlights the difference between each discipline. In games the focus is on maximizing the performance of comparatively simple images that use a limited number of functions under Direct3D or OpenGL. However professionals render more complex scenes and therefore require compatibility to a much broader range of rendering functions. It is this robust driver, software and user support that makes professional products more expensive compared to their gaming counterparts. Where the same manufacture is concerned the hardware can actually be identical between the two markets. As for the Permedia 3 it basically fell between the two stools and was best suited to professionals who liked to play the odd computer game. To date, this has been the last foray into the gaming sector by 3Dlabs, however they are still forging ahead successfully in the workstation market.
At Siggraph99 in August the Bitboys were back. This time with an updated Glaze3D specification that included every feature under the sun, including 9MB of embedded DRAM memory and four pixel pipelines capable of rendering 600Mpixels/sec and 1.2Gtexels/sec. In addition to this, the new chips could be connected in parallel to produce a phenomenal 1.2Gpixel/sec and 2.4Gtexel/sec. Was it merely a coincidence that this specification seemed to mirror many of the forthcoming features from NVIDIA and 3dfx. Would we ever see a Glaze3D chip? Don't hold your breath .
Okay, so I have now cleared the ground for arguably the two most significant releases in 1999. First the Voodoo3 from 3dfx.
Today most producers try to release a product shortly after it is announced. However a few years ago this was not the case and often a considerable period of time would elapse before an announced product would actually hit the streets. The Voodoo3 was a case in point. It was first announced in November 1998 in order to keep to the traditional date for a Voodoo release. The company even demonstrated the new product at Comdex98, however due to the merger with STB and in order to avoid the premature disruption of Banshee sales, the Voodoo3 was withheld until April 1999.
3dfx was one of the first manufactures to re-label their product range in order to target different sectors of the market. So instead of getting a Voodoo3 we were treated to a Voodoo2000, 3000 & 3500 (sound familiar?). Sadly the five-month delay in their release had allowed the competition to catch up and for this reason the new range failed to inspire. It performed well enough, however this performance derived from a legacy of direct industry support via the GLIDE and mini-OpenGL interfaces (APIs). Unfortunately 1999 would witness both the demise of GLIDE and the introduction of competent OpenGL support from the rest of the industry. Strip these away and suddenly the Voodoo3 range was no better than competing cards such as the NVIDIA TNT2 or Matrox G400 released shortly after. In fact the lack of 32-bit color support and AGP texturing led some to conclude that it was 3dfx that needed to catch up. Despite these shortcomings the Voodoo3 sold well, but not quite as well as the TNT2 from NVIDIA.
This unfamiliar situation that 3dfx found themselves in would be further exacerbated by the second significant release in 1999; namely the GeForce256 from NVIDIA. First hinted at in April, but announced in August and demonstrated at ECTS in September, the GeForce256 captured the imagination of the industry by performing the transformation and lighting calculations in hardware (know as Hardware T&L or TnL). Thus the humble graphics chip had been elevated to the status of GPU or Graphical Processing Unit, and although it would be a couple of months before the first retail cards would surface, this innovation alone managed to steal most of the limelight.
3dfx tried to counter this by simultaneously announcing their new parallel T-buffer technology intended for their forthcoming Voodoo4 range. But lack of actual silicon betrayed the fact that 3dfx were nowhere near ready for a release and NVIDIA stole the show.
Ironically, although it was the TnL engine that caught the public imagination, in the end it was the superior fill rate of the GeForce256 that finally established the chip. NVIDIA already had the best selling and one of the best performing chips on the market in the form of their TNT2 chip (equaled by the Voodoo3 but beaten by the G400). The GeForce256 was like two of these chips strapped together. Its ability to render four single-textured pixels in one pass resulted in some impressive scores, sometimes exceeding the performance of a Voodoo3 or G400 by as much as 40%. In December, when DDR memory arrived, this performance gap was further increased to 75%. These performance figures alone held the attention of the market until the following April, when NVIDIA unleashed their second killer chip.
As previously mentioned, August saw S3 announce their Savage2000 chip. In November the first card using this chip was released and came in the form of the Diamond Viper II. The story of the Savage2000 from S3 demonstrates quite clearly the impact of NVIDIA. Everyone had been waiting for the Savage2000 and when it finally arrived it had the second best performance on the street (about equal to a Matrox G400MAX, but slightly less than a GeForce256). However over the following few months the Savage2000 would fade into obscurity. Although troubled TnL support was partly responsible for this, the principle reason was simply bad timing. At any other time the Savage2000 would have succeeded, but at this time everyone was talking about GeForce, Hardware TnL, and the future of 3dfx. Few noticed the Savage2000, and those that did soon forgot about it.
In October Lockheed-Martin closed down REAL3D and sold all remaining assets to Intel who already owned a 20% stake.
At Comdex in November 3dfx finally named their forthcoming products. The cards would be called the Voodoo4 4500, Voodoo5 5000, Voodoo5 5500 & Voodoo5 6000. Each product would utilize one, two or four of their VSA100 chips (Voodoo Scalable Architecture). It was this parallel design that would implement the T-Buffer technology that had been previously announced at ECTS. The principle feature of this new parallel approach was the promise of full and backwards compatible antialiasing at realistic frame rates. Unfortunately this would require more memory than was commonly used and to make the situation worse memory prices had started to climb. This was an ironic twist, because it was the initial fall of memory prices in 1996 that had contributed to the company’s initial success. The development of a modular approach also led to instability problems. This all resulted in the final release being delayed until the following June and then only after a recall. Having to wait five months before releasing their Voodoo3000 range was bad enough, but now having to sit back a further seven months and watch NVIDIA climbed to the top of the hill, must have been totally demoralizing. Could things possibly get any worse for the once golden boys of gaming? You bet your sweat bippy they could!
At least credit has been given via a reference.
