mandelbrot-benchmark

---

Documentation: https://mandelbrot-benchmark.readthedocs.io

Source Code: https://github.com/34j/mandelbrot-benchmark

---

Benchmark Numba, Taichi, Warp, CuPy Kernel, Triton using Mandelbrot set

Results

• 200 iterations max
• $c \in [-2, 1] \times [-1.5, 1.5]$
• JAX version is quite hacky ("vectorized")
• JAX version is omitted because it makes the benchmark unstable (see previous release for the comparison)
• AMD Ryzen 9 3950X + NVIDIA GeForce RTX 4070 SUPER

• The results are almost identical when device = cuda
• The computation time for Numba is quite unstable

Some notes

• In Numba, the type of variables needs to be explicitly specified to use float32 instead of float64 (default), while is not the case for Taichi, Warp.

Installation

Install this via pip (or your favourite package manager):

pipx install mandelbrot-benchmark

Usage

Run the benchmark and plot the results:

mandelbrot-benchmark benchmark
mandelbrot-benchmark plot

Alternatives

• YanagiAyame/python-mandelbrot-benchmark: Which is better, Numba, Taichi, Warp or JAX?
• mandelbrot-on-all-accelerators.ipynb

Contributors ✨

Thanks goes to these wonderful people (emoji key):

This project follows the all-contributors specification. Contributions of any kind welcome!

Credits

This package was created with Copier and the browniebroke/pypackage-template project template.