Comment by bajsejohannes

Pure speculation ahead.

It doesn't have to be better in an absolute sense, but being good enough for a cheaper price, lower power usage, smaller footprint, etc.

I think a lot of floating point calculations could fall into this. For example in neural nets, maybe there are analog versions to calculate the weights, sigmoid function and so on.

And for graphics, you don't really need the exact color value of each pixel. Maybe those could be estimated in analog functions too.