geerlingguy/ai-benchmarks issue #46 — RTX 2080 Ti llama.cpp Vulkan and CUDA llama-bench (build f9ec8858e)

Build f9ec8858e (7474). Vulkan tables on Pi CM5 and the 265K PC, plus a CUDA recompile of the same commit for llama-2-13b Q4_K_M only (pp512/pp2048/tg128, fa 1, mmap 0). llama-2-13b Vulkan rows also used fa 1 and mmap 0 after an out-of-memory failure at the default 4096 prompt. Power is whole-system wall draw.