AI Supercompute Infrastructure
NVIDIA GB200 NVL72
Liquid-Cooled · Rack-Scale Blackwell
AI Supercomputer · NVIDIA · DGX-Class
1.4 ExaFLOPS
FP4 AI performance
peak, per rack
peak, per rack
72
Blackwell GPUs
36
Grace CPUs
2,592
Arm CPU cores
13.4 TB
HBM3e memory (up to)
576 TB/s
HBM bandwidth
130 TB/s
NVLink bandwidth
GPU
CPU
Memory
Fabric
Direct Liquid Cooling
Thermal Architecture · In-Rack CDU
Direct-to-Chip Liquid Cooling
Closed-loop coolant distribution unit
250 kW
Cooling capacity
per CDU, up to
per CDU, up to
~750 L/min
Coolant
flow
flow
1+1
Redundant
pumps
pumps
~120 kW
Rack load
cooled
cooled
GPU
CPU
DIMM
cold plates
Liquid-to-Liquid / Air
Data Platform · VAST Data · DASE
VAST AI Data Platform
Single-rack reference cluster · all-NVMe
8 PB
Raw flash / rack
6× Ceres, up to
6× Ceres, up to
1.375 PB
Per 1RU
enclosure
enclosure
60 GB/s
Per
enclosure
enclosure
99.9999 %
Availability
Global Namespace
GPUDirect Storage
NFS
SMB
S3
Similarity Data Reduction