Liquid-cooled 72-GPU NVLink domain that acts as a single massive GPU for trillion-parameter LLM inference and training.
NVIDIA GB200 NVL72 is a liquid-cooled rack that connects 36 Grace CPUs and 72 Blackwell GPUs so they behave like a single, massive GPU. NVIDIA rates it at up to 30x faster real-time inference for trillion-parameter LLMs than the previous generation.
Each GB200 Grace Blackwell Superchip pairs one Grace CPU with two Blackwell GPUs over NVLink-C2C, and fifth-generation NVLink ties all 72 GPUs together at 130 TB/s for training, inference, HPC and data processing.
Second-generation Transformer Engine with FP4 delivers up to 30x faster LLM inference.
FP8 precision and 1.8 TB/s GPU-to-GPU bandwidth speed up large-scale training.
Up to 25x more performance at the same power than H100 air-cooled infrastructure.
576 TB/s of GPU memory bandwidth for trillion-parameter models.
72 GPUs connected at 130 TB/s remove communication bottlenecks.
Decompression engines make database queries up to 18x faster than CPUs.
| Configuration | 36 Grace CPUs | 72 Blackwell GPUs |
|---|---|
| NVFP4 Tensor Core | 1,440 PFLOPS (sparse) | 720 PFLOPS (dense) |
| FP8/FP6 Tensor Core | 720 PFLOPS |
| INT8 Tensor Core | 720 POPS |
| FP16/BF16 Tensor Core | 360 PFLOPS |
| TF32 Tensor Core | 180 PFLOPS |
| FP32 | 5,760 TFLOPS |
| FP64 / FP64 Tensor Core | 2,880 TFLOPS |
| GPU memory | bandwidth | 13.4 TB HBM3e | 576 TB/s |
| NVLink bandwidth | 130 TB/s |
| CPU cores | 2,592 Arm Neoverse V2 |
| CPU memory | bandwidth | 17 TB LPDDR5X | 14 TB/s |
| Configuration | 1 Grace CPU | 2 Blackwell GPUs |
|---|---|
| NVFP4 Tensor Core | 40 PFLOPS |
| FP8/FP6 Tensor Core | 20 PFLOPS |
| GPU memory | bandwidth | 372 GB HBM3e | 16 TB/s |
| NVLink bandwidth | 3.6 TB/s |
| CPU cores | 72 Arm Neoverse V2 |
| CPU memory | bandwidth | Up to 480 GB LPDDR5X | up to 512 GB/s |
Next-generation rack-scale platform for agentic AI and reasoning, unifying 72 Rubin GPUs and 36 Vera CPUs in a single NVLink domain.
Turnkey, ready-to-deploy enterprise AI infrastructure built on the Vera Rubin platform and managed with NVIDIA Mission Control.
Blackwell Ultra rack-scale system built for AI reasoning, training and data processing at scale.
Tell us your workload and timeline for the NVIDIA GB200 NVL72. We reply within one business day with a configuration, pricing and lead time.