I remember using povray too. Officially it required a 386 pc but it could be recompiled for 286, which is what I did. My PC had a IIT 2c87 coprocessor that had special 4x4 matrix multiply instructions, so of course I modified povray to take advantage of that.
That matmul instruction seemed oddly specific, so I had to google that, and lo and behold, indeed there was such a thing! [0]
Do you remember what kind of speed ups that thing gave you? I wonder if you could play to backport some modern AI algorithms to it and run them at reasonable speeds -- however being a 286 I imagine it'd be quite a challenge.
I no longer remember the numbers but the speed ups were measurable. That FPU had 24 FP registers (as compared to 8 in a regular 287). The best case would be to load a matrix to the registers and then have a bunch of vectors to multiply the stored matrix with. Then this thing could fly. Loading the matrix to the registers was slow though because 287 coprocessors used DMA to communicate with memory.
