Warp experience: we can map computations onto a parallel computer efficiently

Warp is a programmable, systolic array computer developed by Carnegie Mellon and produced by GE. A 10-cell Warp machine can perform 100 million floating-point operations per second (10 MFLOPS). A variety of applications have been mapped onto Warp. The experience has been that the mapping is not a real problem; in fact, usually near-optimal mapping is relatively easy to obtain, and the actual implementation of the mapping on the machine can often be automated. This paper explains why this is the case by examining some computational models which are frequently used on Warp. Carnegie Mellon and Intel are jointly developing a VLSI version of Warp, called iWarp. It is expected that many applications can be efficiently mapped onto low-cost iWarp arrays to achieve an effective computation bandwidth of about one GigaFLOPS.