Warp: a systolic computer in a van
In February 1986 General Electric delivered to Carnegie Mellon the first ten-cell prototype of Warp, a systolic array of linearly connected programmable cells, each capable of 10 million floating-point operations a second, 100 MFLOPS in all. The machine was built for the vision of robot vehicles and rode in the Navlab van.
Why it matters
The systolic array stopped being a diagram: vision programs ran on it daily, and from September 1987 neural network training too. ALVINN's network, which steered the van, was trained on Warp, so training a network here already ran on specialised hardware rather than an ordinary computer.
The project started in 1984 and a two-cell prototype was completed in 1985. General Electric and Honeywell built two wire-wrapped ten-cell prototypes (February and June 1986); General Electric's production PC Warp was delivered from April 1987 at about $350,000. The authors write that Warp is typically several hundred times faster than a VAX 11/780; road following on the prototype took 6,000 ms, a speed-up of 200. Pomerleau, Gusciora, Touretzky and Kung, 1988: one learning trial takes 0.82 ms, about 17 million connections a second; a road network of 105,084 connections trained for 40 epochs on about 50,000 synthetic images in about 3.5 hours of Warp CPU time. While driving, ALVINN by its own 1988 paper ran on the on-board Sun, and so it stayed: in their review The Warp Machine on Navlab (1991) Jill Crisman and Jon Webb write that the trained network was run on a Sun 3 on the van, since applying a network once it was learned took little computation, while the Warp trained it, first in overnight eight-hour runs in the laboratory and from June 1989 on the fly, with a person driving. BYTE in August 1989 gives about 200 milliseconds for the network's forward pass on the Sun-3/160 installed on the van. Carnegie Mellon's report of 1990 gives the dates: colour-based road following on the Warp was demonstrated in December 1988 at 1 m/s, on 16 March 1989 the Navlab set a record of 1.3 m/s under the network's control, and on 6 September 1989 the Warp was taken off the van and replaced by a Sun 4. What the record does not claim: that the Warp papers of 1986 and 1987 name the Navlab; the word does not occur there, and the name comes from the 1988 paper. Nor that ALVINN's network was computed on the Warp while driving: Carnegie Mellon's final report for 1987-1990 says only that the Navlab travels faster than its former 0.5 m/s "using the on-board Warp machine", and two texts by participants name a Sun 3 outright. Pomerleau's report CMU-CS-89-107, which that report cites, was not found.