There is a machine in your computer that is, by most meaningful measures, more powerful than the processor everyone calls the brain. It has more cores. It moves more data. It consumes more power and generates more heat and does more raw mathematical work per second than the CPU has been capable of for over a decade. And it exists because people wanted to play video games.

That's the GPU story in one sentence. Everything else is detail — but the detail is where it gets interesting, because the path from "chip that draws pixels" to "processor that trains artificial intelligence" was not a straight line. It was a forty-year engineering relay race across companies that rose and fell, architectures that were invented and abandoned, and a series of accidents that turned out to be more important than anything anyone planned.

Before There Was a Name for It

In 1981, James Clark — an electrical engineering professor at Stanford — assembled a team of graduate students to study computer graphics. Within two years, they'd built a chip called the Geometry Engine. It was the first VLSI implementation of a geometry pipeline: specialized hardware that could accelerate the inner-loop math needed to display three-dimensional images. Transforming vertices, calculating lighting, removing hidden surfaces — the computational grunt work that made 3D possible. Clark patented it, left Stanford in 1982, and founded Silicon Graphics, Inc.

SGI's first product, the IRIS 1000, shipped in 1983. It was a graphics terminal, not a standalone computer — it plugged into a VAX mainframe. But it could render 3D graphics at speeds that had previously required hardware costing ten times as much. Within a few years, SGI workstations were the standard in Hollywood visual effects, CAD engineering, flight simulation, and scientific visualization. Jurassic Park was made on SGI hardware. So was Terminator 2. The company that would eventually co-found Netscape started by building the machines that taught the world what 3D graphics could look like.

Meanwhile, the PC world was nowhere close. IBM introduced its Color Graphics Adapter in 1981 alongside the original IBM PC — 320×200 resolution, four colors. In 1984, IBM released the Professional Graphics Controller, a three-board monster with its own Intel 8088 processor that could do basic 2D and 3D acceleration. It cost $4,290 — roughly $13,000 in today's money — and almost nobody bought it. The Enhanced Graphics Adapter (EGA) arrived the same year, followed by VGA in 1987, which gave PCs 640×480 at 16 colors and became the baseline standard for the next decade.

Two things were true simultaneously in the late 1980s. SGI was proving that dedicated graphics hardware could produce astonishing results. And the PC — the machine most people actually owned — couldn't draw a smooth diagonal line without visible stairstepping. The gap between what was possible and what was affordable was enormous. Closing that gap would take fifteen years, destroy multiple companies, and produce the most consequential processor architecture of the twenty-first century.

The 2D Era and the Software Stack

Before 3D happened on PCs, 2D happened first — and it mattered more than most histories give it credit for. Windows 3.0 arrived in 1990 and suddenly every PC needed to push pixels fast enough to render a graphical desktop. A market for 2D accelerator chips exploded. Companies like S3, Cirrus Logic, Matrox, and Trident built chips that could accelerate line drawing, rectangle fills, and bitmap transfers in hardware. It was not glamorous work, but it established the principle: offload graphics operations from the CPU to a dedicated chip.

ATI Technologies had been in this game from the start. Founded in 1985 in Markham, Ontario as Array Technology Inc. by K.Y. Ho, Benny Lau, and Lee Ka Lau, ATI built integrated graphics cards for OEMs — IBM and Commodore were early customers. Their VGA Wonder line in the late 1980s, followed by the Mach series in the early 1990s, made them a major player in PC graphics before 3D was even a consumer concept. By 1998, ATI was the market share leader.

The other critical development was the software layer. SGI released OpenGL in June 1992 — a cross-platform graphics API derived from their proprietary IRIS GL library. For the first time, developers could write 3D graphics code that wasn't locked to a specific manufacturer's hardware. OpenGL established the principle that graphics hardware should be programmed through a standardized interface. Then in 1995, Microsoft released DirectX 1.0 — a collection of APIs designed to give Windows developers direct access to hardware for gaming. The stage was set. The hardware was about to follow.

The 3D Revolution

In August 1994, three engineers left Silicon Graphics. Ross Smith, Gary Tarolli, and Scott Sellers had been at SGI during the development of the Nintendo 64's graphics chip, but SGI was going through internal turmoil and didn't want to pursue the gaming market. So they founded 3dfx Interactive in San Jose with a straightforward thesis: bring SGI-quality 3D rendering to a $300 card that plugged into a PC.

They pulled it off. The Voodoo Graphics chipset shipped in late 1996, and it changed everything overnight. It was a 3D-only add-in board — it couldn't do 2D at all, so you needed a separate VGA card connected through a pass-through cable, and some early models had mechanical relays that literally clicked when the 3D chip took over. But when it kicked in, the difference was staggering. Quake went from choppy software rendering to smooth, textured, hardware-accelerated 3D. The Orchid Righteous 3D and Diamond Monster 3D hit shelves at $299, and 3dfx captured an estimated 80–85% of the 3D accelerator market within a year.

Meanwhile, a company founded in January 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem was struggling. NVIDIA's first product, the NV1, launched in 1995 and was a commercial failure — it used quadratic texture mapping instead of the polygon-based approach that the industry was standardizing on. The NV1 even had a Sega Saturn controller port built in, a bet on console-PC convergence that didn't pay off. NVIDIA nearly died. The company pivoted hard, abandoned its proprietary approach, and in 1997 released the RIVA 128 — the first card to combine 2D and 3D acceleration in a single chip at a competitive price. It wasn't better than a Voodoo, but it was cheaper and simpler. NVIDIA was back in the game.

The Voodoo2 followed in 1998, introducing SLI — two Voodoo2 cards working in parallel. NVIDIA answered with the RIVA TNT, then the TNT2. The arms race was on. S3, Matrox, and Rendition tried to keep up and couldn't. The market was consolidating fast around the companies that could sustain the engineering pace required to ship a new architecture every twelve to eighteen months.

The Term Gets Coined

On August 31, 1999, NVIDIA announced the GeForce 256 and called it something nobody had said before: a GPU. Graphics Processing Unit. It was a marketing term — NVIDIA defined it as a single-chip processor with integrated transform, lighting, triangle setup, clipping, and rendering engines. The key innovation was hardware Transform and Lighting — T&L — which moved geometry calculations off the CPU and onto the graphics chip. Previous cards accelerated rasterization — filling in the triangles. The GeForce 256 accelerated the math that created the triangles in the first place.

3Dlabs had actually used the term "GPU" earlier — their Glint Gamma processor in 1997 was marketed as a "geometry processor unit." But NVIDIA trademarked the term and made it stick. The GeForce 256 wasn't just faster than what came before. It was categorically different. It announced that the graphics chip was no longer an accessory to the CPU. It was a processor in its own right, with its own class of work.

ATI responded quickly with the original Radeon in 2000, which they called a VPU — Visual Processing Unit. Nobody used that term. The industry had its name. GPU it was.

3dfx didn't survive to compete. The Voodoo3 shipped in 1999 but couldn't match the GeForce 256. The Voodoo5 followed in 2000 and flopped. On December 15, 2000, 3dfx laid off its entire staff and sold its patents, brand, and remaining assets to NVIDIA for roughly $70 million in cash. The company that had single-handedly created consumer 3D gaming filed for bankruptcy in October 2002. The market was now a two-company race: NVIDIA and ATI.

The Fixed Pipeline Breaks Open

The next inflection point arrived in February 2001 with the GeForce 3. It was the first consumer GPU to support programmable shaders — code that developers could write to control exactly how geometry was transformed and how each pixel's color was computed. Before this, the GPU's rendering pipeline was fixed-function: the hardware did specific things in a specific order and you had no ability to change it. After this, developers could write their own programs that ran on the GPU. The GeForce 3 implemented DirectX 8's Shader Model 1.0, and the fixed-function era ended there.

This seems like a graphics detail. It was the most important architectural decision in the history of computing. Because programmable shaders meant the GPU was no longer a graphics-only device. It was a programmable parallel processor that happened to be aimed at graphics. And a few very observant researchers noticed.

ATI wasn't standing still. The Radeon 9700 Pro in 2002 — the R300 architecture — was the first GPU to support DirectX 9 and Shader Model 2.0, and it was widely considered the best GPU of its generation. ATI and NVIDIA traded blows through the early 2000s, each leapfrogging the other every cycle. The GeForce FX stumbled in 2003. ATI's Radeon X800 was strong. NVIDIA's GeForce 6800 answered in 2004. The competition was brutal, the pace was relentless, and the GPUs kept getting more programmable with every generation.

The Chip Becomes a Computer

In November 2006, NVIDIA released the GeForce 8800 GTX — codenamed G80 — and it represented a fundamental break with everything that had come before. The G80 was the first GPU built on a unified shader architecture. Previous GPUs had separate hardware units for vertex shaders and pixel shaders — different types of programs ran on different parts of the chip. The G80 replaced all of that with 128 identical streaming processors that could execute any type of shader. Vertex, pixel, geometry — the chip didn't care. It just ran programs.

Alongside the G80, NVIDIA released CUDA — Compute Unified Device Architecture. CUDA was a programming framework that let developers write standard C code and run it on the GPU's hundreds of parallel cores. For the first time, you didn't have to pretend your computation was a graphics operation. You could just write a program, target the GPU, and let it rip. The GPU had become a general-purpose parallel computer.

NVIDIA simultaneously launched the Tesla product line — the same G80 chip, stripped of video outputs, packaged for data centers and scientific computing. The GPU that rendered Crysis also ran fluid dynamics simulations and molecular modeling. Same silicon. Different market. This was the pivot point — the moment NVIDIA stopped being just a gaming company and started becoming a compute company.

Meanwhile, ATI's journey took a different turn. In July 2006, AMD acquired ATI Technologies for $5.4 billion. The strategy was integration — combine CPUs and GPUs into a single chip, which AMD would call Fusion (later APU). The Radeon brand survived, the engineering team in Markham kept building, but ATI as an independent company was gone. AMD dropped the ATI name entirely in 2010.

The compute story accelerated. NVIDIA's Fermi architecture arrived in 2010 with proper error-correcting memory and a more capable cache hierarchy — features that mattered to scientific computing, not gaming. Kepler followed in 2012 with a focus on performance per watt. Maxwell in 2014 improved streaming multiprocessor efficiency. Pascal arrived in 2016 with HBM2 memory and NVLink interconnects. Each generation doubled down on the same bet: the GPU is a compute platform, not just a graphics card.

That bet paid off in September 2012 when AlexNet — a deep neural network trained on two GTX 580 GPUs — demolished the ImageNet competition and launched the deep learning revolution. The hardware that NVIDIA built for gamers turned out to be the hardware that AI had been waiting sixty years for. Neural network training is fundamentally a massive parallel matrix multiplication problem, and GPUs were accidentally perfect for it.

Purpose-Built for Intelligence

In 2017, NVIDIA unveiled the Volta architecture and the Tesla V100, and the marketing said everything: "Welcome to the Era of AI." Volta introduced Tensor Cores — dedicated hardware units designed specifically for the matrix math that neural networks require. Previous GPUs could train AI models using their general-purpose CUDA cores. Volta had silicon built for nothing else. The V100 was the first GPU to break 100 teraflops of deep learning performance. It wasn't a graphics card that could do AI on the side. It was an AI processor that happened to share DNA with a graphics card.

The consumer line took its own leap the following year. The Turing architecture in 2018 introduced RT Cores for hardware-accelerated ray tracing alongside Tensor Cores in consumer GeForce cards — the RTX 20 series. Real-time ray tracing had been a rendering goal for decades. Tensor Cores in consumer GPUs enabled AI-assisted graphics techniques like DLSS, where a neural network upscales lower-resolution frames to look like native high-resolution output. The gaming GPU was now running AI inference in real time to make games look better. The two tracks — gaming and AI — had merged on the same silicon.

The Ampere architecture followed in 2020. The A100 data center GPU brought third-generation Tensor Cores, HBM2e memory, and Multi-Instance GPU technology that could partition a single A100 into seven isolated instances. On the consumer side, the GeForce RTX 30 series delivered a massive performance jump. Then came Hopper in 2022 — the H100 GPU, designed from the ground up for training large language models. Its Transformer Engine could dynamically switch between FP8 and FP16 precision depending on the layer and operation type, accelerating the specific architecture that powered GPT-4, Claude, Gemini, and Llama.

The H100 became the most sought-after chip in the world. Data centers couldn't get them fast enough. NVIDIA's market capitalization crossed $1 trillion, then $2 trillion, then $3 trillion. A company founded in 1993 to make PC graphics chips became the most valuable company on Earth. Because the chip that draws pixels turned out to be the chip that trains intelligence.

Forty-five years from the Geometry Engine to the H100. From a Stanford lab trying to make 3D pictures appear faster to silicon that thinks — or at least does the math that makes thinking possible. The GPU didn't start as the most important processor in the world. It started as a shortcut for triangles. Every capability it has today was built on top of a simple, practical insight: some problems are made of millions of small identical operations, and a chip designed to do millions of small identical operations at once will always beat a chip designed to do one complex thing at a time.

That insight turned out to apply to a lot more than pixels.