Geerling llama.cpp Vulkan on rtx3080ti (pc host): tinyllama-11b Q4_K_M pp4096+tg128 6399.16 tok/s

THIRD-PARTY MEASURED (Jeff Geerling, geerlingguy/ai-benchmarks). llama.cpp build d82b7a7c1 (7193), backend Vulkan, default llama-bench flags (ngl 99). Host: Intel Core Ultra 265K, 64GB DDR5, Ubuntu 25.10. Model file size 528.14 MiB, 1.10 B params, Q4_K_M. llama-bench combined pp4096+tg128 figure: one prompt of 4096 tokens followed by 128 generated tokens, reported as a single tok/s. This is neither pure prefill nor pure decode; stored as tps_aggregate with concurrency 1. Whole-system wall power during the run 349.1 W (node scope, not device).

Source issue prints the llama-bench mean and the +/- spread; the stored value is the mean. The Pi-host rows are the same physical GPU on a PCIe-constrained host, so they are not a card ceiling.