Featured image of post GPU Prime Generator

GPU Prime Generator

GPU Prime Generator

Sometimes you just want to see your GPU do math.
GPU Prime Generator is a Python project that uses Numba’s CUDA backend to search for large prime numbers in parallel — no C++, no custom kernels, just high-level Python code running directly on your NVIDIA GPU.

The Idea

Prime generation has always been a fun problem. It’s simple, infinite, and surprisingly good at stress-testing hardware.
Traditionally, you’d write a sieve or a primality test and let your CPU churn away — one number at a time. But GPUs are designed for massive, embarrassingly parallel workloads, which makes them perfect for range testing.

So this project asks a simple question:

How fast can a Python script find primes if we hand the math to the GPU?

How It Works

Under the hood, gpu-prime-generator uses:

  • Numba to JIT-compile Python functions into CUDA kernels.
  • CuPy and NumPy for fast array operations and data handling.
  • Thousands of GPU threads, each testing different numbers at once.

You can specify a starting number and range size, and the tool will fan out those checks across your GPU cores. The results come back almost instantly for mid-sized ranges.

python gpu_prime_generator.py --start 1000000000 --count 10

Example output:

GPU Prime Generator starting at 1,000,000,000
Found 10 primes in 0.42s
Peak GPU threads: 32768

Why Bother?

There are faster ways to find primes — but that’s not the point.
This project exists because it’s cool to watch number theory scale across thousands of CUDA cores.

It’s a great sandbox if you’re curious about:

  • How to map serial math to parallel compute
  • How Numba bridges Python and GPU code
  • Or just how quickly your GPU can brute-force something with minimal overhead

Setup

Requirements:

  • Python 3.10+
  • NVIDIA GPU with CUDA Compute Capability 6.0+
  • numba, numpy, cupy
  • CUDA Toolkit 12+
git clone https://github.com/mgelsinger/gpu-prime-generator.git
cd gpu-prime-generator
pip install -r requirements.txt

Takeaways

Even a small experiment like this highlights how accessible GPU programming has become in Python.
You don’t need C++ or deep CUDA knowledge — just a few decorators and some vectorized logic.

The result? A tool that crunches through primes at ridiculous speeds and doubles as a lightweight benchmark or teaching demo for parallel computing.

Source:
👉 github.com/mgelsinger/gpu-prime-generator

Part of the glsngr.xyz portfolio — open tools and experiments built to explore performance, parallelism, and practical computing.