“SambaNova claims that using NVIDIA for the prefill phase of inference and a SambaNova rack for the decode phase can yield a 2 to 3 times throughput advantage over an all-NVIDIA setup.”