Silicon Graphics (SGI)

Contribution to AI

Before the Geometry Engine there was no chip whose entire job was arithmetic on vertices. The 1982 design by Jim Clark and Marc Hannah pushed the matrix transformations, clipping and coordinate mapping of the graphics pipeline into custom VLSI, and every later graphics processor inherits that shape: a floating-point unit applying the same operation to enormous numbers of values at once, sold at consumer volumes. What was built on top mattered as much. IRIS GL was proprietary and awkward; the 1992 decision to rework it as OpenGL, license it cheaply to competitors and hand the specification to an industry board, meant a single interface to fast 3D across hardware from rival vendors, and the shading languages descended from it became the tools researchers used to smuggle non-graphics arithmetic onto graphics cards. That trick started in-house — Cabral, Cam and Foran used RealityEngine's texture hardware for volume rendering and tomographic reconstruction in 1994, treating texture memory as a numerical array rather than a picture. A decade later people were running neural-network matrix multiplications the same way, through a graphics API, because nothing better existed yet. The workstations themselves were the bench on which 1990s vision, simulation and medical imaging work was done, which is why the machines appear in the acknowledgements of so many papers. The company did not survive the commodity hardware it had made thinkable; its engineers left to found 3dfx and ArtX and staffed the industry that replaced it.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).