NVIDIA GB200 NVL72 – Grace Blackwell Rack-Scale AI Supercomputer

Description

The NVIDIA GB200 NVL72 is a next-generation rack-scale AI computing platform based on the NVIDIA Grace Blackwell architecture. Rather than functioning as an individual graphics card or conventional server, it integrates CPUs, GPUs, NVLink switches, networking, power delivery, and liquid cooling into a unified rack-scale infrastructure designed for extreme AI and HPC workloads.

At the heart of the system are 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell GPUs. Each GB200 Grace Blackwell Superchip combines one Grace CPU with two Blackwell GPUs using the high-bandwidth NVLink-C2C interconnect. NVIDIA specifies up to 3.6TB/s NVLink-C2C bandwidth per GB200 Superchip in the current GB200 NVL72 specifications.

Across the complete NVL72 platform, the 72 Blackwell GPUs provide up to 13.4TB of HBM3e GPU memory and up to 576TB/s of aggregate GPU memory bandwidth. This enormous memory subsystem is designed to keep very large AI models and datasets close to the GPU compute resources while minimizing data movement bottlenecks.

The system’s fifth-generation NVIDIA NVLink fabric provides up to 130TB/s of aggregate NVLink bandwidth, allowing all 72 GPUs to participate in a single large NVLink domain. This architecture is particularly important for trillion-parameter models where conventional PCIe or smaller GPU clusters can introduce significant communication overhead.

The GB200 NVL72 also incorporates Blackwell’s second-generation Transformer Engine, enabling advanced low-precision AI computing including FP4, alongside FP8, FP6, BF16, FP16, TF32, FP32, and FP64 processing. NVIDIA specifies up to 1,440 PFLOPS of sparse NVFP4 Tensor Core performance for the complete NVL72 configuration.

Unlike conventional air-cooled GPU servers, the GB200 NVL72 uses a liquid-cooled rack-scale design. NVIDIA’s reference architecture includes compute trays, NVLink switch trays, power shelves, a bus bar, liquid-cooling manifolds, and an NVLink passive copper cable backplane.

3. Complete Technical Specification

Specification	NVIDIA GB200 NVL72
Manufacturer	NVIDIA
Product Family	NVIDIA Grace Blackwell
Model	GB200 NVL72
System Type	Rack-scale AI / HPC computing platform
CPU Architecture	NVIDIA Grace
CPU Count	36
CPU Architecture	Arm Neoverse V2
Total CPU Cores	2,592
GPU Architecture	NVIDIA Blackwell
GPU Count	72
GPU Configuration	36 GB200 Grace Blackwell Superchips
Blackwell GPUs per Superchip	2
GPU Memory	Up to 13.4TB HBM3e
GPU Memory Bandwidth	Up to 576TB/s aggregate
CPU Memory	Up to 17TB LPDDR5X across the rack
CPU Memory Bandwidth	Up to 14TB/s aggregate
NVLink Generation	5th Generation NVIDIA NVLink
Aggregate NVLink Bandwidth	Up to 130TB/s
NVLink-C2C	Up to 3.6TB/s per GB200 Superchip
NVFP4 Tensor Performance	Up to 1,440 PFLOPS sparse
FP8 / FP6 Tensor Performance	Up to 720 PFLOPS sparse
INT8 Tensor Performance	Up to 720 POPS sparse
FP16 / BF16 Tensor Performance	Up to 360 PFLOPS sparse
TF32 Tensor Performance	Up to 180 PFLOPS sparse
FP32 Performance	Up to 5,760 TFLOPS
FP64 / FP64 Tensor	Up to 2,880 TFLOPS
Transformer Engine	2nd Generation
GPU Interconnect	NVIDIA NVLink + NVLink Switch System
Networking	Qualified high-speed data-center networking, including InfiniBand/Ethernet architectures
Cooling	Liquid cooled
Rack Architecture	NVIDIA NVL72 rack-scale architecture
Compute Trays	18 × 1RU trays in NVIDIA’s current reference configuration
NVLink Switch Trays	9 × 1RU switch trays
GPU Domain	72-GPU unified NVLink domain
Primary Workloads	Generative AI, LLM training, LLM inference, HPC, data analytics
Deployment Class	Enterprise / Data Center / AI Factory

NVIDIA’s published specifications list the complete GB200 NVL72 at 2,592 Arm Neoverse V2 CPU cores, 13.4TB HBM3e, 576TB/s GPU memory bandwidth, 130TB/s NVLink bandwidth, and up to 17TB LPDDR5X CPU memory.

4. Applications

The GB200 NVL72 is designed for extremely large-scale computing environments, including:

Trillion-parameter LLM training
Large Language Model inference
Generative AI
AI model pre-training
AI model fine-tuning
Mixture-of-Experts (MoE) models
Reasoning and agentic AI workloads
High-Performance Computing (HPC)
Scientific computing
Large-scale data analytics
Recommendation systems
Natural language processing
Computer vision
Drug discovery and computational chemistry
Digital twins and simulation
Climate and weather modeling
Enterprise AI factories
Cloud AI infrastructure
Hyperscale AI clusters
Research and supercomputing

NVIDIA specifically positions the GB200 NVL72 for real-time trillion-parameter inference and large-scale training, with the 72-GPU NVLink domain acting as a single massive accelerator fabric.

5. Compatibility

The GB200 NVL72 is a complete rack-scale computing platform, not a standalone GPU or conventional server component.

Platform Compatibility

Designed for:

NVIDIA GB200 NVL72 reference architecture
NVIDIA DGX GB200 infrastructure
Qualified NVIDIA GB200 OEM systems
Enterprise AI data centers
Hyperscale AI infrastructure
NVIDIA-certified AI factories
Large-scale HPC environments
NVLink-based multi-GPU deployments

The NVIDIA reference NVL72 rack architecture uses 18 × 1RU compute trays, 9 × 1RU NVLink switch trays, management/top-of-rack networking, power shelves, and liquid-cooling infrastructure.

Important Compatibility Note

The GB200 NVL72 cannot be installed into a standard 19-inch GPU server as a conventional graphics accelerator. It requires the complete rack-scale architecture, including compatible compute trays, Grace Blackwell modules, NVLink switches, power infrastructure, liquid cooling, firmware, networking, and system management.