We commonly assume a sharp distinction between computer hardware and
software: the digital circuit designer takes logic gates as primitives
to build processors, while the programmer takes processors as
primitives to build applications.  Specialists in one of these domains
seldom cross into the other.

For many years this division of labour has sufficed.  The increased
circuit density that resulted from improved chip manufacturing has
enabled circuit designers to produce faster processors, and everyone
benefited.  That era has now ended, and the current trend is toward
multiprocessors rather than bigger processors.

But there is an alternative way to exploit all those extra transistors
that are now available: you can use them to provide special assistance
to ordinary processors, or even to design circuits that solve problems
directly.

That approach is often unreasonable, because traditional processors
solve many problems efficiently.  There are, however, some
applications where it does make sense to combine circuit design with
programming.  Two examples are specialised functional units and data
parallel processors.  There has been progress in languages and
software tools to support such hybrid systems as well as algorithms
that make the most of their capabilities.  Combining hardware with
software also raises some intriguing new questions in computational
complexity.