GH200 GPU

NVLink-C2C Integration

72 Grace CPU cores and the Hopper GPU communicate at 900 GB/s via NVLink-C2C, delivering 7× higher bandwidth than PCIe Gen5 and enabling coherent memory access across a 576 GB shared memory pool.

Unified Memory Architecture

CPU LPDDR5X (480 GB at 500 GB/s) and GPU HBM3 (96 GB at 4 TB/s) operate as a single addressable memory space, eliminating data copies and accelerating large AI model workloads.

High-Performance Computing

Delivers up to 1 exaFLOPS FP8 AI performance and 3,958 TFLOPS tensor performance, optimized for trillion-parameter LLMs and advanced physics simulations.

Multi-Instance GPU Support

Supports GPU partitioning for multiple concurrent workloads while maintaining full NVLink-C2C connectivity and memory coherence.

Scalable Superchip Design

Multiple GH200 Superchips interconnect via NVLink domains, scaling up to 144 TB of shared GPU memory in DGX GH200 systems for massive parallel processing.

Software Ecosystem Compatibility

Runs NVIDIA AI Enterprise, CUDA, cuDNN, and the HPC SDK natively, supporting all major AI frameworks without requiring code changes.

CPU Subsystem

• Processor: NVIDIA Grace CPU
• CPU Cores: 72 × ARM Neoverse V2 cores
• L1 Cache: 64 KB I-cache + 64 KB D-cache per core
• L2 Cache: 1 MB per core
• L3 Cache: ~117 MB total
• CPU Memory: Up to 480 GB LPDDR5X with ECC
• CPU Memory Bandwidth: ~500–512 GB/s
• PCIe Support: PCIe Gen5 (multiple x16 links)

GPU Subsystem

• GPU Architecture: NVIDIA Hopper
• Streaming Multiprocessors: 132 SMs
• Tensor Cores: 528 Tensor Cores
• GPU Memory: 96 GB HBM3 / up to 144 GB HBM3e
• GPU Memory Bandwidth: ~4 TB/s (HBM3), up to ~4.9 TB/s (HBM3e)

Performance Metrics (Approx.)

• FP64: ~34 TFLOPS
• FP64 Tensor Core: ~67 TFLOPS
• FP32: ~67 TFLOPS
• TF32 Tensor Core: ~494–989 TFLOPS
• BF16 / FP16 Tensor Core: ~990–1,979 TFLOPS
• FP8 Tensor Core: ~1,979–3,958 TFLOPS
• INT8 Tensor Core: ~1,979–3,958 TOPS

Interconnect & Coherent Memory

• NVLink-C2C: Up to 900 GB/s bidirectional coherent CPU–GPU bandwidth
• Unified Memory Model: Shared coherent CPU–GPU address space for reduced latency

Thermal & Power

• Configurable TDP: ~450 W to 1000 W (CPU + GPU + memory)
• Cooling: Air-cooled or liquid-cooled configurations

Form Factor & Software

• Form Factor: GH200 Superchip (integrated CPU + GPU + memory module)
• Software Stack: NVIDIA AI Enterprise, CUDA, TensorRT, full NVIDIA HPC/AI stack

Integrated Superchip Design

NVIDIA GH200 GPU combines Grace CPU with Hopper GPU architecture, forming a unified superchip optimized for AI and high-performance computing workloads.

Ultra-Fast NVLink-C2C

900 GB/s bidirectional CPU–GPU interconnect delivers up to 7× higher bandwidth than PCIe Gen5 for seamless, low-latency data transfer.

Massive Memory Capacity

Up to 96 GB HBM3 GPU memory combined with 480 GB LPDDR5X CPU memory enables efficient processing of trillion-parameter AI models.

Extreme Memory Bandwidth

Up to 4 TB/s HBM3 GPU bandwidth supports rapid data access for large-scale model training and complex scientific simulations.

72 Arm Neoverse Cores

Grace CPU with 72 Arm Neoverse V2 cores delivers significantly higher performance-per-watt compared to traditional x86-based systems.

FP8 Tensor Performance

Up to 3,958 TFLOPS of FP8 tensor core performance accelerates generative AI training and inference for massive language models.

Coherent Memory Model

A single unified CPU–GPU memory space eliminates data copying and reduces latency for memory-intensive AI workloads.

Scalable NVLink Networking

Supports NVLink domain scaling across multiple GH200 systems, enabling exascale AI and HPC supercomputing clusters.