llama.cpp Vulkan Posts 90.5% on Ornith 9B

This is a strong reminder that local inference performance depends heavily on runtime, backend, drivers, and configuration—not just the model or GPU.

• –llama.cpp’s [Vulkan support](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md) gives AMD users a flexible alternative to ROCm-based deployments.
• –The result is compelling for interactive local use, but it reflects one model, GPU, quantization, and benchmark setup.
• –Community testing has also found performance gaps between standalone llama-server and Ollama’s Vulkan path on RX 7000 cards, suggesting integration and version differences matter.
• –Developers should benchmark llama.cpp, Ollama, and ROCm/Vulkan on their own workloads before choosing a runtime.