The Graphics Card Wars, Told by Four Cards in a Kona

Four Cards on a Wet Car Park

In a car park outside Stockholm, a man in red trousers holds up graphics cards one by one and photographs them against the drizzle. A blue Sapphire card with a chrome robot woman printed on the cooler. A red ATI card with a green barcode sticker that reads HD 2900 XT, 512 megabytes. A black Nvidia card with two silver chevrons. A small red Radeon HD 7750 with a fingerprint smudged across the plastic.

To almost anyone, that is a box of electronic waste. To someone who was building computers between 2006 and 2012, it is a timeline. Those four cards, by pure accident, tell the story of the most dramatic six years in the history of graphics hardware. A merger that nearly broke a company. A card so hot and so late it became a joke. A card so good and so cheap it changed what people expected for their money. And, hiding inside the Nvidia story, the quiet beginning of the technology that now runs every large AI model on Earth, including the one narrating this.

So let us go through them in order, as if they were exhibits in a very small museum on the back seat of a Hyundai.

Exhibit One, the Robot Lady

The Sapphire X1950 Pro came out in the autumn of 2006. Sapphire was one of the companies that took chips designed by ATI, the Canadian graphics company based outside Toronto, and built them into cards with their own coolers and their own box art. And box art in those years was its own art form. Every manufacturer had a mascot. Chrome women, armoured knights, dragons, robots, all printed onto the cooler and the box as if you were buying a fantasy novel rather than a circuit board. The robot lady on this card is pure 2006.

The X1950 Pro was a mid-range card, aimed at people who wanted to play the new games of that year without paying for the very top model. It was competent and well liked. It also has a small historical footnote. It was among the first ATI cards that could link to a second card for more performance without an awkward external cable dongle, a feature called CrossFire, ATI's answer to Nvidia's SLI. Two cards in one PC, working together, was the ultimate status symbol of a teenage gaming rig in those years.

But the bigger story around this card is not the card. In July 2006, a few months before it launched, AMD, the processor company that competed with Intel, announced it was buying ATI. The price was around 5.4 billion dollars, an enormous bet. AMD wanted to combine processors and graphics into one company and eventually onto one chip. In the long run that bet paid off spectacularly. In the short run it was chaos. Two very different companies, one in California and one in Canada, trying to merge cultures while their biggest rival in graphics was about to release the most important chip of the decade.

Exhibit Two, the Card That Changed Everything

In November 2006, Nvidia released the GeForce 8800 GTX. The chip inside it was called G80, and it is one of the few pieces of computer hardware that genuinely split history into before and after.

Here is why. Until then, graphics chips had separate specialised units for different jobs. Some units handled the geometry of a scene, the corners of triangles, and other units handled colouring the pixels. The balance between them was fixed in the silicon. If a game needed lots of geometry and little pixel work, half the chip sat idle.

G80 threw that away. It used what is called a unified shader architecture, a large array of identical small processors that could each do any of those jobs, assigned on the fly. Microsoft's new DirectX 10, arriving with Windows Vista, was designed around that idea, and Nvidia got there first with hardware that was ready.

It was not just faster. It was a different kind of machine. Reviewers that winter kept running out of superlatives.

The 8800 GTX was enormous, expensive and power hungry, and it utterly dominated. For roughly a year and a half, nothing ATI or anyone else made could touch it at the top.

But the most important thing about G80 was not games at all. Once you have a chip made of hundreds of small identical processors that can each run any small program, you have something that can do a lot more than draw triangles. In early 2007 Nvidia released CUDA, a way for ordinary programmers to write code that ran directly on those processors for any purpose. Physics simulations. Weather models. Protein folding. Financial calculations. The graphics card, for the first time, became a general purpose parallel computer you could buy in a shop.

For a few years this was a niche interest, mostly for scientists. Then in 2012 a neural network called AlexNet won a famous image recognition competition by a startling margin. It had been trained on two ordinary Nvidia gaming cards using CUDA. That result is widely seen as the moment modern deep learning took off. The whole line from there to today's AI models runs through the decision Nvidia made in 2006 to build G80 the way it did. So the black Nvidia card in the Kona is a distant ancestor of the hardware that trains models like the one you are listening to. Not a direct parent, but family.

Exhibit Three, the Hot One

ATI's answer to G80 was supposed to be a chip called R600, sold as the Radeon HD 2900 XT. The card in the car carries that exact sticker, HD 2900 XT, 512 megabytes, GDDR3, dual DVI.

It was late. Very late. G80 arrived in November 2006, and the 2900 XT did not appear until May 2007, more than six months after. In graphics, half a year is an eternity. And when it finally came, it did not win. It was competitive with Nvidia's second-tier card rather than the flagship, it drew a lot of power, somewhere around 200 watts, and its cooler was loud. Reviewers measured its noise and temperature with a kind of grim fascination.

On paper it was remarkable. It had a 512-bit memory bus, an extraordinarily wide road between the chip and its memory, wider than almost anything before it. It had ambitious features that turned out to be ahead of their time. But games of that era could not use most of that ambition, and real performance lagged behind the specification sheet.

The problem with being early to the right idea is that you pay for it in heat and money before anyone can use it.

The 2900 XT became a byword, a cautionary tale, a card people joked could heat a room. And yet, here is the twist. Retro PC collectors today love it precisely because of that. It is a famous failure, a dramatic chapter, a card with a story. In the small and slightly obsessive world of people who build period-correct gaming PCs from the 2000s, a working 2900 XT is a conversation piece. Some of its ideas, especially around how the shader units were organised, fed directly into ATI's next generation.

And that next generation, the Radeon HD 4000 series in 2008, was a comeback. ATI, now inside AMD, changed strategy. Instead of chasing the biggest, most expensive chip, they built smaller, efficient chips that offered great performance per kronor. The HD 4870 surprised everyone. The lesson of the hot card in the Kona was learned in about a year.

Exhibit Four, the Chevrons

Back to the black card with the two silver chevrons. It looks like the GeForce 8800 GT, and specifically like the version that Apple offered in its Mac Pro from early 2008. If that identification holds, it is a slightly unusual object, a graphics card with Mac firmware, built to work in a cheese grater.

The 8800 GT itself was one of the great value cards of all time. It came out in October 2007, a shrunk-down, cooler, cheaper version of the G80 design built on a newer manufacturing process, and it performed close to the expensive flagship for a fraction of the price. It arrived just weeks before a game called Crysis, a jungle shooter so demanding that "but can it run Crysis" became a running joke for the next decade. The 8800 GT was the card that let ordinary people at least try.

In the Mac world, graphics cards have always been a separate story. Mac cards need special firmware so the machine can show its startup screen before the operating system loads, which meant Mac owners could not just buy any PC card. They had to buy the Apple version, usually at a markup. That made Mac editions scarcer, and today, people keeping old Mac Pros alive specifically look for them. So this card might be the one item in the GPU pile with a genuine little secondhand market, among Mac collectors who want their 2008 machine exactly as it shipped.

Exhibit Five, the Little Red One

The last card is the youngest. The Radeon HD 7750, released in early 2012, here in a version from a brand called VTX3D. It is a small card, with a single fan and no extra power connector, meaning it drew all its power from the motherboard slot, around 75 watts at most. That alone tells you how far the industry had come since the 2900 XT. Five years later, a quiet little card with no power cable was a respectable everyday performer.

It is also historically interesting for a reason most people never noticed. The HD 7000 series introduced an AMD architecture called Graphics Core Next, which they shortened to GCN. GCN was AMD's own move toward the general purpose computing idea Nvidia had pioneered with G80 and CUDA. It made AMD cards much better at non-graphics work. GCN also went on to power both the PlayStation 4 and the Xbox One, which means that for most of the 2010s, almost every console game in the world ran on a close cousin of the chip inside that small red card.

So the little card with the fingerprint is not glamorous. But its descendants were in tens of millions of living rooms.

What a Box of Old Cards Is Good For

Now the honest part. None of these four cards will run modern AI models. Not because of some snobbery about age, but because of memory and software. The biggest of them has 1 gigabyte of memory or less, and modern models want 8, 16, 24 gigabytes or more. The software that runs those models also dropped support for these generations many years ago. These cards were born before the AI boom, even the ones whose great-grandchildren caused it.

What they are good for is everything else. A working 2900 XT, an X1950 Pro and an Apple-edition 8800 GT are each the heart of a period-correct retro gaming PC, and there is a whole hobby built around exactly that. People hunt for Windows XP era machines to play games the way they looked in 2007, on hardware that crackles and whines and smells faintly of hot dust. A weekend LAN party with three such machines would be a perfectly pointless and therefore perfect use of them.

The 8800 GT, if it really is the Mac version, belongs back in a Mac Pro. One of the two cheese graters in the haul might even be the right age for it. And the HD 7750 is the practical one, a spare display card for any machine that needs a screen and nothing more.

There is something nice about this pile. Graphics cards are usually thrown away the moment they are superseded. They are among the most disposable components in computing. Yet these four, collected by a TV studio for whatever reason and abandoned in a box, cover the exact years when the graphics card turned from a toy for gamers into the engine of a new industry. The robot lady, the hot one, the chevrons, and the little red one. A merger, a failure, a bargain, and an architecture that ended up in every console.

Not bad for a box of e-waste.