News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
GH200 Grace Hopper Sup

GH200 Grace Hopper Superchip

Combined Arm CPU and Hopper GPU for AI and HPC workloads

The GH200 Grace Hopper Superchip merges an Arm-based Grace CPU with a Hopper GPU using NVLink-C2C. Designed by Nvidia for large-scale AI and high-performance computing, it enables coherent memory access between processor types to accelerate data-intensive tasks.

Share on WhatsApp

  • Combines Arm Grace CPU and Hopper GPU
  • NVLink-C2C offers 900 GB/s coherent bandwidth
  • Supports up to 480GB LPDDR5X CPU memory
  • Configurable power range from 450W to 1000W
  • Enables coherent memory access for AI workloads

What it is

The GH200 Grace Hopper Superchip is a combined processor unit from Nvidia. It integrates a Grace central processing unit based on Arm architecture with a Hopper graphics processing unit. The two chips connect via NVLink-C2C, a high-speed interface that allows them to share memory directly. This design targets giant-scale artificial intelligence and high-performance computing applications where moving large datasets quickly is critical. By removing the need to copy data between separate CPU and GPU memory pools, the system aims to boost efficiency for scientific research and complex modelling tasks.

How it came about

Nvidia developed this superchip to address bottlenecks in analytics and query processing caused by the slowing of Moore’s law. Traditional systems often struggle with data transfer limits between processors. The company combined its Grace CPU, which uses Arm Neoverse V2 cores, with the Hopper GPU architecture. This pairing was designed from the ground up to support modern AI workloads. The product line includes configurations that may use either H100-style or H200-style GPU memory, reflecting an evolution in available hardware components during its development phase.

Inside the chip

The chip features 72 Arm Neoverse V2 cores for the CPU portion. These cores handle preprocessing tasks efficiently. The GPU side uses Hopper architecture with either 96GB of HBM3 or 144GB of HBM3e memory. A key feature is the coherent memory model, where CPU and GPU threads access a single page table. This setup eliminates redundant data copying. The CPU includes LPDDR5X memory soldered onto the package, available in 120GB, 240GB, or 480GB sizes. The NVLink-C2C interface provides 900 gigabytes per second of bandwidth, which is significantly faster than standard PCIe Gen5 connections.

How it was made

Nvidia packages the CPU and GPU together in a single superchip unit. The LPDDR5X memory modules are soldered directly onto the package, with eight on top and eight below the CPU die. This on-package design reduces power usage and improves performance by shortening signal paths. The GPU memory stacks vary, with six stacks for the 96GB version. Power delivery comes directly to the package rather than through the motherboard. The total power draw can range from 450 watts to 1000 watts, depending on the configuration and cooling method used.

Where it went

This superchip powers systems for large language models, graph neural networks, and scientific simulations. It supports retrieval-augmented generation by accelerating vector searches and embedding processes. Data centres use it for AI training and inference, as well as high-performance computing tasks like molecular chemistry and fraud detection. The chip runs Nvidia software stacks including AI Enterprise and the HPC SDK. Systems based on this hardware often pair it with BlueField-3 data processing units for networking and storage offload.

Versions and family

There are multiple configurations under the GH200 name. CPU memory options include 120GB, 240GB, and 480GB of LPDDR5X, with varying bandwidths. The GPU component may be a 96GB HBM3 version or a 144GB HBM3e version, the latter resembling the H200 chip. Some 144GB units report 141GB usable capacity due to yield reservations. Power limits are configurable, allowing operation between 450W and 1000W. The NVL2 variant connects two GH200 chips via NVLink to provide up to 1.2TB of fast memory and higher bandwidth for single-server workloads.

Why it matters

The superchip aims to deliver energy-efficient AI processing at scale, with combined systems reaching hundreds of exaflops. It addresses the shift in bottlenecks from compute to memory bandwidth in modern analytics. By integrating CPU and GPU memory, it offers speed improvements for data-heavy tasks compared to traditional dual-socket CPU setups. The design influences how data centres approach memory-intensive workloads, offering a alternative to discrete CPU and GPU architectures for specific high-performance needs.

Problems and reception

Industry observers note that the GH200 naming convention covers a wide range of specifications, which can cause confusion. The term applies to units with different memory sizes, bandwidths, and power limits. Some users have clarified that not all GH200 units contain the newer H200-style GPU memory, as some use H100-style components. Despite this complexity, testers report significant performance gains in clustering and data processing tasks compared to previous generation hardware.

Questions people ask about GH200 Grace Hopper Superchip

What is the GH200 Superchip?

It is a combined processor from Nvidia that integrates an Arm-based Grace CPU and a Hopper GPU. They connect via NVLink-C2C to share memory coherently, designed for large-scale AI and high-performance computing tasks.

How much memory does the GH200 have?

The CPU side has 120GB, 240GB, or 480GB of LPDDR5X memory. The GPU side has either 96GB of HBM3 or 144GB of HBM3e memory. Total capacity depends on the specific configuration purchased.

Is the GH200 the same as the H200 GPU?

No. The GH200 is a superchip containing both a CPU and a GPU. The GPU part may be similar to an H100 or an H200, but the GH200 name refers to the entire combined package, not just the graphics processor.

What is the power consumption of the GH200?

The chip can operate between 450 watts and 1000 watts. The exact power draw depends on the workload and configuration, with higher power settings typically requiring liquid cooling for optimal performance.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on GH200 Grace Hopper Superchip yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/grace-hopper-superchip/, servethehome.com/a-quick-introduction-to-the-nvidia-gh200-aka-grace-ho. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    Photo credits on this page

    The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.