In SGI parlance, a pipe (see Graphics pipeline) is akin to a PCI Video Card connected to other cards that handle Geometry and/or Rasterization in a "pipeline" (hence the name and the relevant nature of OpenGL), somewhat similar to 3Dfx's Voodoo Graphics card (though many considered the latter hackish and not at all a true pipe). Nowadays a pipe is solely on a single board and on a single chip (GPU). More to the point, a single SGI pipeline (e.g. InfiniteReality4) could be configured to have up to 8 Video Outputs, each with its own Raster Manager (or 4 with additional Texture memory and a Geometry manager). SGI's Implementation of Xinerama would transparently handle OpenGL on all display outputs (including multiple pipes), but usually offloaded all OpenGL calls to a blanket OpenGL API called transperant OpenGL(like running Xdmx without Chromium).
Programming for a single display on multiple pipes was always possible in OpenGL using threads (and subsequently in Open Inventor and OpenGL Performer), making them "multipipe aware", but this proved challenging for scientists and arquitects who cringed on C/C++ multithreading. SGI decided to lighten the burden by developing a package that would encompass a halfway optimized solution for applications programmed for only one pipe via Multipipe.
It was part one of a three tiered approach to escalate applications towards multipipe awareness. The second part was Multipipe SDK which provided a simple set of libraries that made an OpenGL application multipipe aware. The third was moving on to the multipipe aware OpenGL Performer. The latest versions of Multipipe include Xdmx as a better suited solution for multiple displays. Ultimately this solution would be superseded by working Xdmx, Xinerama and Chromium in tandem on linux graphics clusters (Whilst programming multiple pipes in OpenGL would be facilitated by Equalizer or by a variety of SceneGraph Libraries).