SRC are talking about the following characteristics for a field-programmable floating-point device that they intend to put in one of the variants of their SRC-7 reconfigurable computer:
Hard floating point units with field-programmable interconnect
Each FP unit also performs 53 or 24 bit integer ops with 106 or 48 bit results
Selectable 150 MHz or 300 MHz operation
I've heard that these units will have single-cycle latency. I don't imagine that these FPGAs will be cheap as there must be considerable NRE costs that won't be recouped so easily as they would be for a Xilinx chip. However, they open up a whole new world of possibilities for scientific computing, making double-precision much more viable than it has been previously.
Is this the device reconfigurable computing has been waiting for?