Toshiba
Did you know there were 16-bit 3D accelerators? Toshiba made one in 1990, smooth shading being the killer feature, it was named Gouraud Shading Processor. Instead of targeted 20 MHz, the TC8512 chip had to settle for 12 or 16 MHz, resulting in 8 MPx/s fillrate at best. But more could be combined for double or quadruple rates. Or you could use a pair to render 24-bit color pixel with 8-bit depth value every second cycle. Division Ltd was going to use one chip per display in their VR. The following GSP2 should have been a 32-bit 33 MHz texture mapper but to my knowledge was not used in games.
Fujitsu
Their biggest claim to fame can be the graphics co-processors used by Sega arcade machines. In 1992 groundbreaking polygon count of Virtua Racing was realized with TGP MB86233 16 MFlops FPU made for matrix operations.
In 1995 Fujitsu was among the several chosen ones with integrated texturing accelerator: the Advanced Graphics Processor (yes, abbreviated as AGP). At 60 MHz it measured 15 MPx/s fillrate with point sampled textures. Fujitsu also had their TGPx4 Geometry Processor available for triangle setup and even separate chip for Z-sorting. All of these could run in parallel, thus enabling many multi-chip configurations able to achieve top-class performance for that year. The TGPx4 was so flexible that it could be configurable for TnL as well. Fujitsu thus had the most complete pipeline, from the front-end to blending operations at the end. But when it came to gaming their 3D chips were pretty much only used in arcades.
PC entry could happen in 1998 when in team with IBM they were about to equip Rendition's V2200 with TnL front-end, but that project was cancelled.
Yamaha
By 1995 two affordable chips using Rendering Polygon Accelerator architecture are offered. It was a classical rasterizer with hardware texture mapping. First the YGV611 with more expensive VRAM prohibiting it in consumer space. But the later YGV612 with regular DRAM was still considered strong enough for 640x480. Western Digital created Tasmania 3D card with an initial price tag of $250 and two Renderware games in the box. Although real retail product, since I cannot get a hold of her, it is listed here.
The joys of texturing without perspective correction.So unceremoniously ended years of 3D research and Yamaha returned to their VDPs.
Samsung
In 1996 as BRender support shows, Samsung was about to enter the gaming market with Black Dragon. The KS82C614 chip, integrating everything from triangle setup to texture mapping, aimed at consoles and CAD. Little is known about it other than supporting 64-bit PCI and WRAM memory. There was also KS82C615 and since Samsung was a partner in the development of Talisman API, clearly PC was a target as well. Then the trace of 3d acceleration is lost until it emerges a decade later in the form of floating-point coprocessor inside their mobile SOCs.
Brooktree
The BTV chipset arrived on PC by the end of 1995 with a lot of promise, but 3D acceleration was to be realized only with the next product, the BTV2120. However, the company struggled with the support of BTV features alone and was sold at the beginning of 1997 without ever showing their 3D solution.
S-MOS
The company entered the 3D market with the SPC1500, a single-chip 150 MFlop geometry processor targeting NT workstations. S-MOS had a pixel processor at work to be combined with the SPC1500 for a full accelerator. But the evolution led into a very affordable product: Pix.
The Pix was clearly targeting gaming, as an advertisement in September 1996 shows. Possibly the lowest cost solution, textures are to be stored in system memory and the small card has nothing on it other than the PIX chip. Still, the declared fillrate is 66 MPixels/s high. With three open Api to be supported out of the box, where is the catch? The Pix finishes work with rasterization, the pixel color writes have to be calculated and written by the host and then dumped to the video card's frame buffer. Since depth values are precalculated it is clear why the boards have no memory. This partial 3D acceleration did not make it to the PC. I can see the window of opportunity closing when others have more complete solutions selling already and CPUs are quickly getting faster. Make no mistake, the PIX was no vaporware and was offered at least for Apple platform.
Stellar
S-MOS collaborated with RSSI- a team around the Pixel Squirt architecture. The people behind it gave it one more try under new company Stellar to bring their first 3d chip VelaTX to market in 1999. Just like CRT displays, VelaTX would process pixels in the order of beam. It would be another solution eliminating Z-buffer by hidden surface removal. Ace in the sleeve should now be 2.5 MB of 125 MHz eMDRAM, connected by 512-bit bus, exclusive for textures. Compression should have made the size sufficient, but additional external memory was an option as well. Thanks to the EDRAM properties Stellar tackled demands of even the most messy (z-wise) frames, enabling very stable framerates. But again, no partner would deliver VelaTX for PC, Stellar's portfolio was only interesting for set-top box products.
Oak Technology
The saddest story is probably that of Warp 5, a product canceled when boxes are already shipping. It would be the second TBDR architecture next to PowerVR. Oak went a step above them by implementing the PC industry's first multi-sampled anti-aliasing with such confidence it cannot be turned off. On-chip Z-Buffer uses 24-bit floating-point precision. Since at least few boxes got outside, we have an idea of the card being clearly slower than Voodoo, but winning quality-wise. Oak never tried again after 1997.
Tseng Labs
In 1996 ET6000 was making strides with ridiculous high bandwidth for its class. What a solid platform for your own 3D accelerator: the ET6300. Integrated full triangle setup engine, sophisticated texture filtering, and Direct3D support. Supposed to arrive in early 1997, with AGP version later. In July Tseng is still waiting for the first samples. By the end of the year, the company is still bleeding money without any 3D chip seen. Buyout by ATi happened, leading some to the fantastic conclusion of Rage 128 being the ET6300 in disguise. AFAIK there is no reason to believe this.
Philips
The potential of VLIW processor handling OpenGL was examined from the very first Trimedia chip. The solution was aimed at accelerating many tasks of the new multimedia possibilities, quite like Chromatic. Of course, the chip could handle the whole pipeline, the question is how fast. The triangle setup was done with ease, but bilinear texture filter was more demanding. The concept was proven by GLQuake benchmarks clearly beating software solutions. Either way the option to promote Trimedia as 3d accelerator was there since 1997. Philips instead decided to wait for TM-4 with dedicated graphics hardware. By that time specialized 3D accelerators fully succeeded and the Trimedia features were scaled back to video and audio processing.
ARK Logic
Advanced Rendering Kernels, another company of industry veterans from silicon valley, enjoyed some 2D success with efficient 1000PV, but following series 2000 had a much harder time. In May 1998 ARK announced entry into the 3d accelerator market with 8100 Tiger 3D. This engine promised single-cycle perspective corrected, z-buffered, edge anti-aliased, trilinearly filtered pixel processing. At 100 MHz and aggressively priced the Tiger looked alive and dangerous. Frame buffer had a 64-bit memory interface, but textures had their own dedicated 32-bit bus. With a fast triangle setup, ready for true colors, it seemed to have it all. But no cards were ever for sale. ARK had one more announcement early in 1999: Cougar 3D. The press release mentions Winbench 3D, so the silicon was probably really brought up. Although clocked 10 % faster than Tiger 3D, Cougar performed on par with Rage Pro Turbo and the Intel i740. That is low-end for 1999, explaining ARK's aggressive pricing. Yet again, no cards were seen in the market. After these failures, ARK Logic ceased activity.
Bitboys Oy
For years the most mysterious team in the field, teasing consumers with the best gaming chips coming to the market very, very soon. Since they never materialized, Bitboys can still be remembered by many as a vaporware company. Once upon a time in the demoscene, there was a Finnish Future Crew pushing hardware to the max with their 3D programs. Two of the crew formed Bitboys in 1991. After two more years of coding, hardware research commenced. Who else should know what accelerator is needed for the most spectacular graphics than masters of demoscene? In 1995 they joined forces with others, notably Tritech to market the product: Pyramid3D. Initial chip TR25201 was bug-ridden, but the 25202 without geometry engine got to a stage to be able to run a lot of Direct3D apps. However, the development continued well into 1996 mainly to incorporate 2D as well. Meanwhile, Tritech lost all money, and Bitboys were left without a partner. Thus the most feature-rich gaming 3D accelerator of the time did not happen. That does not mean the Pyramid did not leave its mark. Microsoft used the implementation of environmental bump mapping in DirectX 6. It would just take one more year to get hardware supporting it. Despite that, Bitboys began work on the next architecture again with TnL but also embedded memory enabling groundbreaking anti-aliasing performance. Only to have their memory partner go bankrupt right when the silicon taped out. By that time Bitboys changed its target market to mobile one and years later finally scored with the first pixel shading chips featuring serious FSAA. They just refused not being the best.
Gigapixel
Founded in 1997, Gigapixel offered their design GP-1 in October 1999. I am not saying chip because similar to Imagination technologies, they used the IP licensing model. Another similarity is Tile-based rendering. Rendering by tiles small enough to fit into local cache where occluded parts are removed drastically reduced memory bandwidth and fillrate requirements. Gigapixel's renderer is traditional though, no infinite planes. With one peculiar advantage- FSAA without performance penalty! Just in time when anti-aliasing techniques recieved more attention. The GP-1 demonstration board would beat the best competitors in Quake 3 framerates, while having a very low gate count. And yet it seems like there was only one customer interested. Were AIBs afraid of a new player with new technology in times of rapid consolidation? That would not hurt should the one customer be as big as - Microsoft.
However, Microsoft rejected Gigapixel from X-Box graphics contenders. Right after the board approved buy-out by 3dfx. Why was 3dfx interested? Struggling with their long cycles of own chips they could gain drop-in product and expand their market because GigaPixel prepared low power design as well. With their core blocks, the minimal configuration would process only 16-bit color and depth values. Double the block to double the precision. Gigapixel develops GP-3 chip with powerful front-end, but then 3dfx collapses, and Nvidia acquires everything, including the intriguing intellectual property of Gigapixel.