➀ NVIDIA's Blackwell Ultra GB300 NVL72 achieves 45% higher inference throughput than GB200 in MLPerf's DeepSeek R1 test, leveraging upgraded tensor cores and software optimizations;

➀ The system uses NVFP4 quantization to shrink model sizes and NVLink 1.8TBps interconnect for sharding Llama 3.1 405B across 72 GPUs;

➂ NVIDIA positions GB300 as a cost-efficient solution for 'AI factories,' with shipments beginning this month to drive tokenized data center workflows.