GPU-Resident Top-K Kernel Speeds RAG Retrieval 8.6x

A custom CUDA Top-K kernel keeps similarity search resident on GPU memory, eliminating PCIe round-trips in agentic RAG. The architecture achieves an 8.6x speedup over optimized CPU baselines on a 7-year-old GTX 1080. Only the query embedding and K results cross the PCIe bus.