For as long as computers had drawn pictures, there were two ways to make an image. There was the way movies did it — simulating actual rays of light, letting them bounce, scatter, and reflect through a scene until every shadow and gleam fell exactly where physics said it should. And there was the way games did it — a bag of clever shortcuts called rasterization, faking reflections with mirrored copies of rooms and painting shadows on ahead of time, because the honest method was hopelessly, laughably too slow.
The honest method had a name: ray tracing. Pixar's render farms chewed on single frames for hours. A video game had about sixteen milliseconds. For decades the gap between those two numbers was treated as a law of nature. Then, in 2018, NVIDIA decided to break it.
The physics Hollywood could afford and games couldn't
Ray tracing is beautiful because it is simple. To find the colour of a pixel, you shoot a ray from the camera into the scene and follow what light actually does — it reflects off the chrome of a car, refracts through a glass of water, bounces off a red wall and tints everything nearby a faint pink. Do this for millions of rays and you get images indistinguishable from photographs. This is why every summer blockbuster used it.
The problem was arithmetic. Each ray has to be tested against thousands of triangles to see what it hits, and a single frame needs millions of rays, each spawning more. Doing that on a normal GPU meant seconds or minutes per frame — fine for a film studio, absurd for a game running sixty times a second. Rasterization stayed king not because it was better, but because it was the only thing fast enough. Everyone assumed real-time ray tracing was, in Jensen Huang's words, “still a decade away.”
Turing's three-engine gamble
NVIDIA's answer was not to make ray tracing fast in software, but to build silicon that did nothing else. The new architecture was called Turing, and its trick was that a single chip now carried three different kinds of engine working on the same frame. The familiar CUDA cores still rasterized and shaded geometry the old fast way. New RT cores did one job obsessively well: tracing rays through a scene and finding what they hit. And Tensor cores — borrowed straight from NVIDIA's AI datacenter chips — ran neural networks to clean up the grainy, half-finished ray-traced image and reconstruct the missing detail, a feature christened DLSS, Deep Learning Super Sampling.
The groundwork had been laid at GDC in March 2018, when Microsoft announced DirectX Raytracing (DXR) — a standard way for games to ask for rays — and NVIDIA revealed its RTX technology to accelerate it. The hardware arrived that summer. First came the professional Quadro RTX cards at SIGGRAPH on August 13. A week later, on August 20 at Gamescom in Cologne, Huang walked on stage to unveil the consumer GeForce RTX 20 series: the RTX 2070, 2080, and the monstrous 2080 Ti.
The flagship chip, TU102, was a beast of engineering — 18.6 billion transistors packed onto a 754 mm² die built on TSMC's 12nm process, fed by fast new GDDR6 memory. It was also expensive. The RTX 2080 Ti launched at $999, and $1,199 for NVIDIA's own Founders Edition — a startling jump that made the top card cost as much as a decent laptop. On stage, demoing reflections shimmering across a scene, Huang kept repeating a phrase that would follow him for years: “It just works.”
A rough launch, a long payoff
It did not, at first, entirely just work. When the cards shipped in September 2018, there were almost no games that used ray tracing at all. The showcase titles trickled in over the following months — reflections in Battlefield V, shadows in Shadow of the Tomb Raider, full lighting in Metro Exodus — and turning the feature on could cut frame rates in half. DLSS, the AI upscaler meant to buy back that performance, launched blurry and underwhelming. Reviewers were blunt: gorgeous technology, thin software, eye-watering price. Buyers were paying today for a promise about tomorrow.
But the promise was real, and the bet was strategic. By welding ray tracing to a hardware standard and shipping it in millions of gaming cards, NVIDIA forced the entire industry to move. Game engines rebuilt their lighting around it. A refined DLSS 2.0 arrived in 2020 and quietly became one of the most valuable features in PC gaming, turning AI upscaling from a gimmick into a necessity. Within a few years every competitor — AMD, Intel, even the next generation of consoles — had scrambled to add ray tracing of their own. NVIDIA had not just built a faster card; it had redrawn the map and made everyone else follow.
The Tensor cores hinted at something bigger, too. The same AI silicon that denoised a game frame was the same silicon reshaping NVIDIA's other business entirely — the one filling datacenters, not living rooms. To feed that business, Jensen was about to go shopping for the one thing GPUs alone couldn't provide: the wiring that ties thousands of them together.
Next time: Buying the Plumbing — the $6.9 billion Mellanox deal and how NVIDIA quietly bought its way into the heart of the data center.
Comments
Post a Comment