Nvidia gives the graphics processing unit its lasting name

What happened

Earlier accelerator cards drew triangles that the computer's processor had already positioned and lit. This one took over both steps, calculating where objects sat in space and how light fell on them before rasterizing the result. In practice that meant a scene could contain far more geometry without the processor becoming the bottleneck, and the balance of work inside a computer shifted accordingly. Calling the chip a processing unit rather than a controller or an accelerator was a marketing decision, and it was also an accurate description of what had changed.

Competition in the graphics business at the end of the decade was fierce and short-lived. Several firms with strong products failed to keep pace with a release cadence that turned over roughly twice a year, and consolidation followed quickly. Nvidia bought the assets of its most prominent rival within about eighteen months of this launch. The pattern of shipping new parts on a rapid schedule, each substantially faster than the last, was established during these years and has continued with remarkable consistency ever since.

The architecture kept absorbing functions. Fixed lighting calculations gave way within a couple of years to programmable shading, which let developers write small programs that ran on the card itself, and those programs turned out to be useful for work having nothing to do with drawing pictures. Scientific computing, financial modeling and eventually the training of machine learning systems all ended up running on hardware designed for games, while console makers began licensing graphics architectures from the same handful of companies.

The world at the time

Consumer 3D acceleration was three years old and crowded by 1999. A dozen companies sold chips, standards were still settling, and buyers had to navigate competing proprietary interfaces. The main processor still did the geometry work in every consumer system, which placed a hard ceiling on scene complexity no matter how quickly the card could draw pixels.

What it changed

Programmable shading arrived within two years and made the hardware general enough to be used far outside graphics. Microsoft licensed a version of the architecture for its first console. The graphics business narrowed to a small number of firms, and the parallel processors descended from this line became the foundation of modern machine learning, an outcome nobody selling game cards in 1999 had predicted.