NVIDIA GH200 Grace Hopper Superchip – 96GB HBM3 AI & HPC Accelerator

Description

The NVIDIA GH200 Grace Hopper Superchip is a purpose-built accelerated computing platform that integrates the NVIDIA Grace CPU and NVIDIA Hopper GPU into a single coherent superchip.

The Grace CPU incorporates 72 Arm Neoverse V2 cores with Armv9-A architecture and SVE2 support. It is paired directly with the Hopper GPU through NVIDIA NVLink-C2C, providing up to 900GB/s of coherent bidirectional CPU-to-GPU bandwidth. This substantially reduces the communication bottleneck associated with conventional PCIe-connected CPU/GPU architectures.

The original GH200 configuration provides 96GB of HBM3 GPU memory with up to approximately 4TB/s-class memory bandwidth, alongside up to 480GB of LPDDR5X CPU memory. The coherent memory architecture allows GPU workloads to access CPU-attached memory directly, creating a much larger addressable memory space for applications that exceed the GPU’s local HBM capacity.

The Hopper GPU portion contains 132 SMs and supports fourth-generation Tensor Cores, Transformer Engine technology, DPX instructions, and CUDA acceleration. This makes the GH200 particularly effective for large AI models, scientific simulations, recommender systems, vector databases, and other workloads where both compute performance and memory capacity are critical.

A major advantage of GH200 is its coherent unified memory model. CPU and GPU threads can access shared memory without requiring traditional explicit data movement between separate CPU and GPU memory pools. NVIDIA describes this architecture as enabling the GPU to address both its HBM3 and Grace CPU LPDDR5X memory through NVLink-C2C.

The GH200 is not a conventional standalone PCIe graphics card. It is a server-class superchip/module intended for specialized accelerated computing platforms and large-scale AI/HPC infrastructure.

3. Complete Technical Specification

Specification	Details
Manufacturer	NVIDIA
Product	Grace Hopper Superchip
Model	GH200
Platform	NVIDIA Grace Hopper
CPU Architecture	Armv9-A
CPU Core Architecture	NVIDIA Grace / Arm Neoverse V2
CPU Cores	72 cores
CPU SIMD	4 × 128-bit SVE2 per core
CPU L1 Cache	64KB instruction + 64KB data per core
CPU L2 Cache	1MB per core
CPU L3 Cache	117MB
CPU Memory	Up to 480GB LPDDR5X
CPU Memory Type	LPDDR5X with ECC
CPU Memory Bandwidth	Up to approximately 500GB/s
GPU Architecture	NVIDIA Hopper
GPU SMs	132
GPU Memory	96GB
GPU Memory Type	HBM3
GPU Memory Bandwidth	Approximately 4TB/s-class
GPU Compute Capability	9.0
Tensor Cores	4th Generation
GPU L2 Cache	Up to 60MB in the original Hopper architecture
CPU-GPU Interconnect	NVIDIA NVLink-C2C
NVLink-C2C Bandwidth	900GB/s bidirectional
PCIe	PCIe Gen5
PCIe Connectivity	Up to PCIe Gen5 x16-class links, platform dependent
Memory Architecture	Coherent CPU-GPU unified address space
ECC	Supported
Form Factor	Server superchip / module
Primary Deployment	Data center / AI / HPC
Typical Workloads	AI, ML, LLM, HPC, scientific computing, analytics

NVIDIA documentation identifies the original GH200 as a 72-core Grace CPU with up to 480GB LPDDR5X and a Hopper GPU with 96GB HBM3.

4. Applications

The GH200 Grace Hopper Superchip is designed for:

Large Language Models (LLMs)
Generative AI
AI model training
AI inference
Deep learning
Transformer-based workloads
Recommendation systems
Vector databases
Natural language processing
Computer vision
High-performance computing (HPC)
Scientific simulations
Engineering simulations
Computational fluid dynamics
Molecular dynamics
Pharmaceutical research
Financial modeling
Seismic and geophysical processing
Large-scale data analytics
Digital twins
GPU-accelerated databases
AI supercomputing

5. Compatibility

The GH200 Grace Hopper Superchip is intended for specialized NVIDIA Grace Hopper server platforms, rather than conventional desktop or standard PCIe GPU systems.

Compatible deployment environments include:

NVIDIA Grace Hopper accelerated computing platforms
GH200-based AI servers
NVIDIA DGX GH200-class infrastructure
GH200 NVLink-connected systems
Qualified OEM accelerated-computing platforms

NVIDIA’s DGX GH200 architecture combines GH200 superchips with the NVLink Switch System, allowing GPUs across the system to operate within a large shared NVLink addressable memory environment.