Inference speed benchmark for llama-3-8b with half precision (FP16) and quantization technique.