the hardware for low-precision matrix multiplications in ML is mostly independent of the hardware that’s typically utilized for vertex and fragment shaders in computer graphics. It really only gets used for graphics if using DLSS. It’s not used at all for (integer) computations in cryptomining.

Are they for high-precision scientific simulations because it’s possible to multiply matrices on them? These cards are probably quite shit at double-precision matrix multiplications.

Yes and they make Quadros for those :)