bryanoliveira / cellular-automata

A cellular automata program built with C++, OpenGL, CUDA and OpenMP.

Hi there 👋 I'm Bryan

Passionate computer scientist, thrilled by innovation and amazed by the human brain. Interested in high-performance computing and working towards artificial general intelligence through deep reinforcement learning research and machine learning engineering. Voracious meta-learner in the free time.

• 🔭 I’m currently working on machine learning engineering and reinforcement learning research applied to debt collection.
• 🌱 I'm always learning, as I find that constantly expanding my horizons helps to boost my creativity.
• 💬 Ask me about AI, how we can draw inspiration from the brain, and philosophical questions about it.
• ⚡ Fun fact: I once wrote a program that could simulate Life within Life itself:

---

cellular-automata's Issues

Please clarify: This program is memory bandwidth limited, nothing else

I think it would help the next person learning about this stuff to mention that in your readme.

That's why you don't see much speedup when adding more cores.

I measured/calculated 20% memory bandwidth efficiency for the GPU implementation and 25-50% for the CPU implementation.

None of your evaluations or observations mention memory bandwidth at all.

Also, using curand is what makes your implementation not support even a 30k * 30k = 900 MB grid. sizeof(curandState) = 48 bytes...
If you get rid of curand and use any other decent hash, you can max out the size of the grid.