A Japanese AI company called Preferred Networks has a mature open-source library for NumPy/SciPy calculations on GPUs.

It's called CuPy 🚀.

For massive datasets, it is often enough to replace a single line: import cupy as cp
The same array operations can run on CUDA up to 100 times faster.

What it can do:
🛠 Highly compatible with existing NumPy and SciPy code
📝 Dramatically reduces the need to rewrite code or learn new syntax
💻 Supports not only NVIDIA CUDA but also AMD ROCm architectures

Keep in mind:
→ Only faster for massive arrays; small datasets will run slower due to CPU-to-GPU data transfer lag
→ Strictly bound by your physical GPU VRAM limits (can cause out-of-memory errors).
→ Covers most major math functions, but does not replicate 100% of NumPy/SciPy modules.

The project is completely open-source and battle-tested since 2015 📂: https://github.com/cupy/cupy