Replacing hardware that thinks (especially about parallelism) with a very smart compiler

There is a tradeoff that touches many areas of computing: doing work at run time, usually in the hardware, vs. doing work in software at compile time. This tradeoff exists in some very visible and other not as visible places in any computer system. One surprising place in which it has a large effect is in the specification of fine-grained parallelism. The author examines that tradeoff, particularly in the light of VLIW architectures. VLIWs are very parallel new machines that take the software alternative to the extreme. They are the flip side of dataflow architectures, finding the same individual operation level of parallelism, but dataflow's opposite along the axis represented by the above tradeoff. In the past year, surprisingly successful general-purpose computers have been built from VLIW technology. >