It's now 2006.

Most of what you read here is from personal memory, not news sources, so some of the dates may not be completely accurate. Also, I jump around a bit, because the chronological order doesn't necessarily make for the best narrative.

Ten years ago, I was living in Corvallis, Oregon. Back then, I was a freelance writer, mostly writing reviews and technology features for Computer Gaming World. That year, two different Bay Area companies trekked up to Central Oregon to pay me a visit. One was Rendition, the other Nvidia. Nvidia showed off the NV1, a "multimedia accelerator" that had hardware support for 3D graphics, audio, and video. Rendition had samples of its Vérité V1000, which was a combination 2D/3D graphics chip.

Meanwhile, 3dfx (3Dfx back then) had started shipping the original Voodoo card towards the end of 1995. The Voodoo card was 3D only, so required a separate 2D card to run Windows and non-gaming applications. 3dfx generated considerable mindshare because of its effective developer relations activities. Danny Sanchez, then of Orchid, dragged me into the 3Dfx developer's seminar at the 1995 Computer Game Developer's Conference. At that same conference, Mike Weksler and I crashed the Rendition suite. At that point, 3Dfx had working silicon; Rendition only had emulation loops running. It was an era of accelerated rasterization with triangle setup, bilinear filtering, and minimal MIP-mapping thrown in.

Most of the hardware in development had started gestating before Microsoft bought a UK company known as Reality Lab. Reality Lab's 3D API became Direct3D 1.0. At the 1996 GDC, ATI handed out versions of the original Rage3D graphics card, with a whopping 2MB of memory on board. We also saw S3 jump into the fray with its Virge 3D chip. Both the Virge and Rage3D were often referred to derisively as "3D decelerators", because running in 3D mode was sometimes slower than using software rendering—though the image quality was better. Continued... After the Direct3D announcement, a bunch of companies jumped into the fray. At one point, analyst Peter Glaskowsky counted no fewer than 56 semiconductor firms developing 3D graphics chips. The architectures ranged from the fairly traditional to fairly exotic stuff, like the PowerVR's tile-based rendering system and IDT, which was developing a graphics processor with hundreds of lightweight processors.

But the darling of gamers was a system with a Rendition PCI card and two Voodoo2's in SLI mode. Despite Microsoft's efforts, it was a time when proprietary APIs created a great deal of confusion among developers and consumers. 3Dfx had the biggest mind share, with its right-to-the-metal Glide API. By 1999, though, it was Glide, Direct3D, and OpenGL.

Meanwhile, Nvidia almost imploded after anemic sales of the NV1. The problem with the NV1 was that it used quads instead of triangles. But Microsoft's Direct3D was triangle-based, so getting DirectX to work with the NV1 was an onerous chore.

Around then, I began to understand the real impact of hardware reviews. I've always treated reviews as a way of communicating the real behavior of products to end users—the people who have to shell out real money for a piece of technology that may or may not do the job. Hardware companies treat product reviews as ways of leveraging additional sales. One marketing staffer back in 1998 told me that a Computer Gaming World review of a graphics card was worth plus or minus $20 million to their bottom line, depending on whether the review was positive or negative. But I still maintain that reviews are for readers and people who are looking to buy products, not the companies that sell said products.

At the 1996 Game Developer Conference, Matrox showed up with its Mystique accelerator. The Matrox people spent about a half-hour extolling the value of various filtering algorithms (such as bilinear), really pumping us up. Then they said, "But we don't do any of that."

The idea was interesting—adding sophisticated features adversely affected performance. But Matrox's presentation of that idea fell a little flat.

At any rate, by 1997, Dave Salvator and I decided it was time to take a stand, and decided to use only benchmarks using APIs that were industry standards, whether they were de facto or determined by standards bodies. That meant DirectX and OpenGL. It also meant not using GLIDE, which was still a major hardware-specific API. This eventually resulted in a somewhat tense meeting with 3Dfx at the 1998 E3 Expo. My memory is a little fuzzy, but I seem to recall that Andy Keane, Scott Sellers, Greg Ballard, and a couple of others from 3Dfx were present. They wanted to convince us that we should use Glide for benchmarking. The point was well taken, as anyone using a 3Dfx product would be using the Glide layer in the game—and often getting better performance. Dave and I believed that Glide was starting to look a little old compared with Direct3D and OpenGL, and that it was holding back game development. Both sides remained unconvinced at the end, although we had what might be called a "healthy exchange of views."

Of course, 3Dfx wasn't the only consumer 3D pioneer that failed to track the industry. Both 3Dfx and Rendition weren't keen on shipping AGP products. Back then, AGP in its 1x and 2x incarnations only had a fairly slight performance advantage over PCI. No one really believed that storing textures in system memory and fetching them as needed was viable: The bandwidth wasn't there. Even in today's world of PCIe x16 slots, the bandwidth isn't good enough for certain game titles. In fact, Intel shipped the i740, a graphics chip that relied heavily on storing textures in system memory. Performance was somewhat weak, to say the least, despite specs that looked good on paper. Continued... In 1997, Brian Burke—then at STB, showed up on my doorstep. By 1997, I was back in Sunnyvale, and Brian arrived with a RIVA 128 board. The RIVA 128 was developed by Nvidia, which had gone into quiet mode for a couple of years after NV1 tanked. Rumors had surfaced about the RIVA 128, suggesting that it was really fast (for the time), but no one believed it. Everyone believed that Nvidia was dead.

The RIVA 128 made effective use of AGP but didn't rely on it in the same way that the Intel i740 did. And it was fast—much faster than anything else around. As soon as benchmarks began to be published, 3Dfx shifted from talking about performance to talking about image quality. The RIVA 128 did have some image quality issues, mostly revolving around the auto-MIP-mapping algorithms. But the RIVA 128—plus 3Dfx's inability to develop new products that were better than old ones—started the long downward spiral. Eventually, 3Dfx bought STB, which baffled almost everyone, shut off 3rd parties from selling 3Dfx products, and became a vertical supplier of both graphics chips and cards. Eventually, they shipped the V5500 and, almost, the V6000.

Meanwhile, Rendition was also imploding. The company shipped the V2000 and, eventually, the V2200. There was even an AGP version.

The V2200 AGP was really a PCI part bolted onto an AGP card. In fact, one card manufacturer built a card with an AGP connector on one edge and a PCI connector on the opposite edge. You could just flip the card over for your interface du jour.

But Rendition couldn't get subsequent products out the door. While the RIVA 128 gave Nvidia valuable revenue and breathing space, Rendition's delivery for the V4000 stretched out. Eventually, the company was acquired by Micron, who was interested in developing core logic. To date, nothing much has come of that effort, either.

Matrox is still around, but developing cards mostly for vertical segments. Nvidia, in a clever business move, acquired the assets of 3dfx (note the lower case "d"), but not the company itself, and so dodged any liabilities that might have come with a corporate acquisition. Nvidia was on a roll, too, coming out with the GeForce 256, the first card that brought parts of the geometry pipeline into consumer hardware. Prior to the GeForce 256, all 3D cards for desktop PCs were actually rendering accelerators.

Meanwhile, ATI had been developing a series of increasingly mediocre 3D parts. It was beginning to look like an all-Nvidia world. Then ATI stunned the industry by shipping the Radeon 8500, which offered performance nearly on parity with Nvidia's best GPUs. Then ATI managed to beat Nvidia to the punch with the first DirectX 9.0 GPU, the Radeon 9700, which caught Nvidia completely off guard. But in a "it seemed like a good idea at the time" move, Nvidia responded with the NV30, aka the GeForce FX 5800 a few months later. But the 5800 was noisy, late and still ceded the performance crown to the 9700 and subsequent ATI models. Unbowed, Nvidia developed the 6800 line, which used a more conventional architecture.

So now it's really a horserace between two companies, with S3, XGI, and Matrox really just nipping at their heels for the leftovers. The industry has gone from 56 players down to two and some change. Now that's some industry consolidation.

This Week on ExtremeTech

Jason also takes a look at a new gaming mouse from, of all things, Creative Labs. Also, Victor Loh sizes up an Intel 945G Express motherboard from Foxconn and compares high-capacity hard drives from Seagate and Western Digital. Meanwhile, Loyd Case ponders what's needed to build a Windows Vista–capable system.

Be sure to drop by the forums and post your thoughts on the week's happenings.