That was my point, read carefully the question

Any sort of strategy is useless if you use either Python or ASM as long as any sort of higher-level op like SHA / RIPEMD is the actual bottleneck.

But if we only operate on EC then it's another story, this is where kangaroo / rho / bsgs comes into play, and we can start optimizing the operations that are the bottleneck at THAT level. It's one thing to do a few thousand mul/s in Python and another to do batched additions (with a single inversion) on a GPU, to reach giga adds/second. Millions of times faster for identical results. And ofcourse it's still not enough even at that level, so...

Something to think about on 256-bit modular math:
- a single inversion is ~60 times slower than a multiplication (GCD algo, not a simple operation at all)
- a multiplication is ~2 times slower than a squaring
- a squaring is ~3 times slower than an addition, ~2 slower than a normalization

A simple point addition requires one inversion, 2 multiplications, a squaring, 2 normalizations, and 6 additions.
A simple point multiplication requires a bunch of point additions (in truth, it's somewhat faster bcz of Jacobian shortcut, but still a lot of multiplications).

So stuff like doing point-scalar multiplication is a bad joke, compared to the problem of optimizing the primitives