GB300 NVL72 achieves high token generation speeds using TileRT on DeepSeek v4.1 Flash

The GB300 NVL72 system achieves up to 1,725 tokens per second per user on DeepSeek v4.1 Flash using TP24 with TileRT MegaKernels. This performance is reported to be twice as fast as general-purpose decode engines. Access to the hardware was provided by Ian Buck for the TileRT project. Open article →