“Using NVIDIA GPUs for the prefill phase and SambaNova racks for the decode phase of inference can provide a 2 to 3 times throughput advantage at the same cost.”