I've done some comparisons before, and the conclusion is that it really depends on the computation task. Generally speaking, Numba on CPU is very similar to Taichi in terms of concepts and performance. Numba on GPU is more like writing CUDA, while Taichi allows you to use the same code for CPU & GPU.

You can refer to this repo for more detailed benchmark results: https://github.com/taichi-dev/taichi_benchmark/tree/main/suites

Numba is rubbish because they don't tell you how many unreasonable errors you will encounter using it (a lot) for anything that is not a very trivial function. That's despite reading its docs in detail. It is overrated.