Cray Research

Contribution to AI

A machine built to do one thing — floating-point arithmetic on long arrays, as fast as physics permitted — ended up defining what a computing resource is. The Cray-1 delivered to Los Alamos in 1976 held its vectors in registers rather than streaming them from memory, which made it fast on the short loops real programs actually contain instead of only on synthetic ones; the curved cabinet existed to keep wires short, because signal travel time had become the limiting term in the design. The compiler mattered as much as the hardware. Ordinary Fortran loops were vectorised automatically, so scientists got the throughput without writing anything unfamiliar, and the idea that a compiler should restructure a program to suit parallel silicon starts in earnest here. The X-MP then made multiple processors sharing memory a normal thing to buy, shifting the argument from faster arithmetic units to more of them. Chess supplied the visible AI result: Cray Blitz ran on this hardware and won the world computer chess championship in 1983, holding it in 1986, one of the earliest demonstrations that raw search throughput could substitute for cleverness. When the National Science Foundation established its supercomputer centres in 1985 it filled them with these machines and networked them, and remote access to scarce compute, rationed by proposal, became an academic habit. The deeper legacy is the assumption that a research programme's ceiling is set by the arithmetic it can afford.

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