geerlingguy/ai-benchmarks issue #45 — RTX 4070 Ti llama.cpp Vulkan llama-bench (build 933414c0b)

Same two-host protocol. Build 933414c0b (7293), Vulkan, ngl 99. gpt-oss-20b on the 265K PC re-run with -n 128 -p 512 and -fa 1 -mmp 0 after ErrorOutOfDeviceMemory; those rows are not mixed with the default-flag rows. Power is whole-system wall draw.