GPU

The GPU emerged because 3D graphics had become too complex for CPUs to handle alone. By the late 1990s, games like Quake demanded real-time rendering of thousands of polygons, lighting calculations, and texture mapping. Graphics accelerators handled some of these tasks, but transform and lighting (T&L)—the mathematical operations that determine how 3D objects appear from different angles under various lights—still burdened the CPU.

Nvidia's GeForce 256, announced August 31, 1999, was "the world's first GPU." The marketing term was new; the term "GPU" had appeared earlier (3DLabs used "geometry processor unit" in 1997), but Nvidia defined it precisely: "a single-chip processor with integrated transform, lighting, triangle setup/clipping, and rendering engines that is capable of processing a minimum of 10 million polygons per second." By moving T&L from the CPU to dedicated hardware, the GeForce 256 freed the processor for game logic and physics.

The adjacent possible had aligned through a decade of graphics hardware evolution. Video display controllers had given way to 2D accelerators, then to 3D accelerators like 3Dfx's Voodoo. TSMC's improving process technology—the GeForce 256 used 220nm—allowed 17 million transistors on a single chip. DirectX 7.0 provided standardized APIs for developers. And the gaming market had grown large enough to justify the R&D investment in specialized silicon.

The name "GeForce" stood for "Geometry Force," emphasizing the chip's geometric transformation capabilities. The NV10 graphics processor ran at 120 MHz with 32MB of SDR video memory. It outsold competitors by offering performance that made previously impossible visuals—real-time shadows, complex reflections—achievable on consumer hardware.

The cascade from GPU to AI was unforeseen. For a decade, GPUs evolved along their designed path: faster polygon rendering, programmable shaders, higher resolutions. But around 2007, researchers noticed that GPUs' massively parallel architecture could accelerate tasks beyond graphics. Neural network training, which requires multiplying millions of matrices, mapped perfectly onto GPU hardware. When deep learning exploded after 2012, GPUs became the infrastructure of AI.

Nvidia pivoted to embrace this accidental capability. CUDA (2006) provided programming tools for general-purpose GPU computing. By 2024, Nvidia's data center revenue—AI training and inference—exceeded its gaming revenue. The company's market capitalization reached $1 trillion, built on chips designed for games but essential for AI.

Path dependence locked Nvidia into dominance. The CUDA ecosystem created switching costs; code written for Nvidia GPUs required rewriting for competitors. AMD and Intel offered alternatives, but Nvidia's installed base and software tools maintained a commanding lead. The GeForce 256's 17 million transistors grew to 76 billion in the RTX 4090—a 4,400x increase in 25 years.

By 2026, the GPU has evolved from gaming accelerator to the fundamental hardware of artificial intelligence. The chip Jensen Huang's team designed to render polygons faster now trains the models reshaping every industry. The accidental universality of parallel processing made a graphics chip the engine of the AI era.