You could use several ADSP (Analog Devices) chips as a coprocessor.

Other solutions :

Infineon (ex Siemens) 80C537 includes a 16 bit, fast, fixed point math
coprocessor and a shifter. You have all ingredients to emulate
a reasonable fast floating point system.

Triscend E5 chip includes a FPGA, where you could implement a true
floating point unit. FPGA macro cells are RAM programmed, so you
could reprogram indefinitely your chip.