Overview
My friend and I built a high-performance Wordle solver using NVIDIA GPUs and parallel processing techniques. Our solution takes an information-theoretic approach to evaluating guesses, where we compute the Shannon entropy for each word in a preset dictionary. We build CUDA kernels to concurrently process data and compute the expected information for each guess. With over 11k words to search, we measured a (12x) speedup over a serial implementation of the algorithm.
You can read more here.
Technical Details
- Implements parallel word elimination using CUDA kernels
- Optimized for GPU memory access patterns
- Coalesced memory access
- Caching highly accessed word patterns in shared memory
- Ran experiments on an NVIDIA Quadro RTX 6000