Using AMD MI300X for High-Throughput, Low-Cost LLM Inference