“SambaNova's SM40 and SM50 platforms are 5 to 10 times faster on the decode phase of inference compared to an NVIDIA GPU.”