NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference

NVIDIA blog on GB200 NVL72. Key facts: 72 Blackwell GPUs in single NVLink domain, 130 TB/s total NVLink bandwidth, 1.8 TB/s bidirectional per GPU (14x PCIe Gen5), 30 TB unified memory. GB200 NVL36 and NVL72 configurations. 9 NVLink switch trays, each with 144 NVLink ports at 100 GB/s. 5th-gen NVLink scales to 576 GPUs in a single domain with over 1 PB/s total bandwidth. GB200 compute tray: 2 Grace CPUs + 4 Blackwell GPUs, 80 PFLOPS AI performance, 1.7 TB fast memory. 30x faster real-time LLM inference vs H100 (1.8T parameter GPT-MoE, TTL=50ms, FTL=5s, 32768 input/1024 output). 4x faster training. 25x energy efficiency.