A different floating point hack makes exp() easier to compute in hardware (and consequently tanh). You cast the input to an int and take the first 2 bits of what would be the mantissa. LUT[Index] and LUT[Index+1] from your 5-entry table are used to either lerp or poly approx. the function, with the remaining mantissa bits to help.
A different approach, refining the square root based sigmoid with a polynomial, is in my blog post "a few of my favorite sigmoids" [1]. I'm not sure which is faster without benchmarking, but I'm pretty sure its worst case error is better than any of the fast approximations.<p>[1]: <a href="https://raphlinus.github.io/audio/2018/09/05/sigmoid.html" rel="nofollow">https://raphlinus.github.io/audio/2018/09/05/sigmoid.html</a>
There’s an analysis of the Schraudolph approximation of the exponential function (along with an improvement upon it) that someone might find interesting at <a href="https://typ.dev/attention#affine-cast" rel="nofollow">https://typ.dev/attention#affine-cast</a>
Looks interesting. Should start with a definition of the Hyperbolic Tangent. It is only about 2/3 of the way that the definition occurs in a discussion of computing exp(x).