XiongjieDai GPU-Benchmarks-on-LLM-Inference — llama.cpp cuBLAS on RunPod: LLaMA-3 8B/70B Q4_K_M + F16, generate + prompt-eval tok/s across NVIDIA + Apple parts

llama.cpp built with cuBLAS (make LLAMA_CUBLAS=1, May-2024 era). ./main -ngl 10000 full offload; llama-bench -p/-n 512/1024/4096/8192. Ubuntu 22.04, CUDA 12.1.1, single-GPU rows are per-request batch-1; multi-GPU rows marked '* N'. Numbers verbatim from README tables.