I finally got my hands on a couple of MM57409 "Super Number Cruncher" chips with 1985 date codes. I think it was probably introduced around 1982. It's a NMOS part with the same concept (though not compatible with) the earlier PMOS MM57109 "Number Oriented Processor", which was introduced in 1977.
In both cases, National took their existing 4-bit masked-ROM microcontroller designs (MM5799 PMOS, COP440 NMOS), and the floating point code they'd already written for calculator chips, and turned them into math coprocessors. The MM57109 also had some support for acting as a floating point general purpose processor. The COP4xx has a test mode, so I should be able to dump the ROM of the MM57409. The MM5799 almost certainly had a test mode as well, but I haven't uncovered any documentation for it, so the ROM might have to be dumped optically. The MM57109 uses an 8-digit mantissa, and a divide takes an average of 78ms, and worst-case 223ms. The MM57409 has a 12-digit mantissa, and worst-case dvidie time is 66ms, with no average stated. Both also have a reasonably full complement of functions that would be found on a typical scientific calculator, including transcendentals. Both are quite slow compared to contemporary math coprocessors such as the AMD Am9511A (second-source by Intel as the 8231A), and insanely slow compared to the 8087. I'm tempted to wire up one of these (either kind) to an Apple II, and hack Applesoft to use it as a floating point decelerator. Back in the day, a few companies sold Am9511A cards for the Apple II. Of course, since Applesoft uses binary floating point but these National Semiconductor chips use BCD, the necessary conversion code running on the 6502 would make it even slower. Perhaps Atari BASIC would be a better choice as it used BCD.