AMD lifted the lid on its prize graphics chip to attack new markets with the R580. In the past, graphics cards were designed to be black boxes that conformed to an operating system's view of a peripheral, with a device driver providing standard application programming interfaces (APIs) such as DirectX or OpenGL. Hitting the bare metal was rare (and usually unnecessary) as the complexity of the API grew.
This isolation isn't a problem for the initial duty of graphics cards. But programmers became envious as the power of these cards expanded to match or exceed the performance of the main processor. A GPU could do things that a CPU wouldn't even attempt, or at least get done this year.
The idea of coprocessors isn't new. What is new is increased interest in using standard accelerators for a range of topics. It could bring multiple card or rack systems back into vogue in many application areas that have moved toward single-board computers.
Chips like Azul Systems' 48-core Vega 2 Java chip would be interesting on an x16 PCI Express card. A wide range of search and processing algorithms can benefit from optimized as well as parallel architectures, as indicated by the popularity of DSPs. Developers' appetites for performance are as insatiable as their need for more storage.