Infrastructure atlas
NVIDIA/Rack-scale system

GB200 NVL72

A liquid-cooled rack-scale domain that treats 72 Blackwell GPUs and 36 Grace CPUs as one tightly coupled AI system.

ASICrack scaleBlackwellNVLinkliquid cooling
GB200 NVL72 hardware

01 / GPU count

72 B200

One NVLink domain

02 / Unified memory

30 TB

Across the rack

03 / GPU fabric

130 TB/s

NVLink Switch System

Precision & peak compute

PrecisionPublished peakBasis
NVFP4 Tensor1,440 PFLOPS72-GPU rack · sparse (720 dense)
FP8 / FP6 Tensor720 PFLOPS72-GPU rack · sparse
INT8 Tensor720 POPS72-GPU rack · sparse
FP16 / BF16 Tensor360 PFLOPS72-GPU rack · sparse
FP325,760 TFLOPS72-GPU rack

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

NVIDIA GB200 NVL72

Technical specification

Rack composition
CPU
36 Grace CPUs
Accelerators
72 Blackwell GPUs
Cooling
Liquid cooled
Topology
72-GPU fifth-generation NVLink domain
Scale-out & software
GPU-to-GPU
1.8 TB/s NVLink per GPU
Network
InfiniBand and Ethernet options
Software
NVIDIA AI Enterprise, Magnum IO, CUDA
Target
Trillion-parameter training and real-time inference