|
DISCUSS
2026-03-19 | Independent Researcher, Shenzhen, China | 2603.19163
M=5
P=6
I=7
cuGenOpt is a GPU-accelerated metaheuristic framework that uses a 'one block evolves one solution' CUDA architecture and JIT compilation to solve combinatorial optimization problems. The results are rigorously backed by hardware benchmarks, showing it matches OR-Tools on small instances and vastly outperforms MIP solvers, though it struggles with large-scale VRP (>200 nodes) due to memory limits. WHAT WE LEARNED: The framework allows user-defined search operators (CUDA snippets) to be JIT-compiled and evaluated massively in parallel on the GPU. For us, this is highly actionable: we could use cuGenOpt as the high-throughput fitness evaluation backend for AlgoEvo, taking our LLM-generated ALNS operators, compiling them via their JIT pipeline, and evaluating them across thousands of instances in seconds to solve our scalability bottlenecks.
|