RDU

SambaNova combined four Nvidia H200 GPUs with 16 SN50 RDU chips to achieve 763 tokens per second on MiniMax M2.7 at a 10,000-token context, separating prefill and decode stages to accelerate inference…