By 2006, NVIDIA had won the argument it started at that Denny's booth thirteen years earlier. The GPU was real, it was fast, and it lived in tens of millions of gaming PCs. But inside the company a stranger idea was taking hold — one that had almost nothing to do with games. What if the graphics chip, that dense slab of silicon built to shade pixels, could be talked into computing anything? Not triangles and textures, but weather models, molecular dynamics, financial risk, the raw linear algebra that underpins half of science.
It sounds obvious now. In 2006 it sounded like a distraction. That was the bet.
The problem with a chip that only drew pictures
A graphics processor was, even then, a monster of parallel arithmetic. To paint a screen sixty times a second it had to run the same simple math across millions of pixels at once. That is a wildly different shape of machine from a CPU, which fields a few very clever cores that chew through instructions one after another. The picture below is the whole story in one frame: a handful of heavyweight CPU cores on the left, and on the right the GeForce 8800 GTX's small army of identical workers.
Researchers had already noticed. Through the early 2000s a small, obsessive community practiced a dark art called GPGPU — general-purpose computing on graphics hardware. To run a physics simulation they had to disguise it as a rendering job: pack their numbers into textures, pretend they were colors, and trick the graphics pipeline into doing the math as a side effect of drawing. It worked, barely, and it was miserable. You needed to be fluent in OpenGL or Direct3D just to add two arrays together.
One of those frustrated researchers was a Stanford PhD student named Ian Buck. In 2004 he built a language called Brook that let you write ordinary-looking code and quietly compiled it down onto the GPU, hiding the graphics trickery underneath. It was a prototype, but it proved the point. NVIDIA noticed, and hired him.
Building a chip that meant it
The insight NVIDIA committed to was that the disguise had to go. Instead of bolting compute onto a graphics pipeline, they would design a processor that was genuinely programmable from the ground up — and give it a way to be addressed in plain C, no textures required. Ian Buck and hardware architect John Nickolls led the effort. They called it CUDA, for Compute Unified Device Architecture, a name NVIDIA would later quietly stop spelling out.
The hardware arrived first. On November 8, 2006, NVIDIA launched the GeForce 8800 GTX, built on a new architecture named Tesla. It was the company's first unified shader design: instead of separate fixed units for different graphics tasks, it had 128 identical, general-purpose cores that could be pointed at anything. Under the hood sat roughly 681 million transistors — an enormous chip for its day. To gamers it was simply the fastest card on Earth. To the CUDA team it was a 128-way parallel computer that happened to also render Crysis.
The software followed in February 2007, when NVIDIA released the CUDA toolkit as a public beta. For the first time, a scientist could sit down, write a few lines of C with a couple of new keywords, and have it run across all 128 cores. No pretending numbers were pixels. No graphics degree required. The dark art of GPGPU became, almost overnight, an afternoon's work.
The expensive part nobody could see
Here is where the bet got costly. NVIDIA did not build one special compute card and call it done. Jensen Huang's decision was to make every GPU the company shipped CUDA-capable — every GeForce in every gaming rig, whether or not its owner ever ran a single scientific kernel. That meant spending transistors, die area, heat and money on a feature the vast majority of customers would never touch, year after year, on the theory that a market which did not yet exist would eventually show up.
Wall Street was not charmed. For years CUDA looked like a line item that generated more cost than revenue, and analysts said so. Huang kept funding it anyway, framing it as building a road before there were cars — if the hardware was everywhere, the software ecosystem and the researchers would come. It was the same instinct that had bet the company on the RIVA 128, scaled up and stretched across a decade of quarterly earnings calls.
What NVIDIA had quietly done was turn its entire product line into an installed base of programmable supercomputers, sitting in dorm rooms and labs and offices, waiting. They had built the machine. Now they needed the world to find a reason to use it — and the first place it caught fire was not a trading floor or a games studio, but the research bench.
Next: GPUs break out of the desktop and into the world's laboratories — a supercomputer, suddenly, in every lab.
Comments
Post a Comment