AMD ROCm vs. Vulkan performance for Lemonade local AI server with Llama.cpp
The benchmark results don’t point to one backend winning everywhere. Across tests on Strix Halo and Radeon AI PRO R9700 systems, Vulkan often delivered better token throughput, while ROCm sometimes reached the first token sooner; individual model and task results varied. The tests used Lemonade 2026.39.1 with its bundled Llama.cpp build, so treat the findings as a snapshot of those stacks and systems.