There’s a matrix multiplication in the middle of a lot of popular things, right now, and it’s an operation which frequently tolerates remarkably low precision.

In the SIMD world precision usually bottoms out at 8-bit. You don’t save much by trying to get less precise than that, except for marginal savings in how often you have to mitigate overflows.

But what if you slice one matrix into bitplanes and then you replace multiplication with a hardware accelerated parallel conditional addition? You should be able to fit eight conditional adds into something like the same space as an 8-bit multiplier (but with fewer result bits), and you can take eight times as many input rows at once, to compensate for the fact that you’re only doing one eighth of a multiply.

To get the same precision takes eight iterations, over each of the eight planes of the input, with shifts and adds to merge the results, but you gain the opportunity to quit after fewer iterations for a proportional saving in overall run time.

It seems to fit neatly into a SIMD architecture, but I’ll have to go into that detail a bit later on…