Grace CPU Superchip
Dual-chip Arm module for AI data centres
The Grace CPU Superchip is a server module by NVIDIA that links two Grace CPUs. It offers 144 Arm cores and high memory bandwidth for efficient AI and HPC workloads.
- Contains 144 Arm Neoverse V2 cores
- Uses up to 960GB of LPDDR5X memory
- Delivers up to 1TB/s memory bandwidth
- Operates within a 500W power envelope
- Connects two chips via 900 GB/s NVLink
What it is
The Grace CPU Superchip is a high-performance computing module designed by Nvidia for modern data centres. It combines two Grace CPU chips into a single unit using a fast interconnect. This design targets artificial intelligence, data analytics, and scientific simulations. The module aims to deliver high performance while using less energy than traditional server processors. It integrates a large number of processor cores and substantial memory capacity in a compact form factor.
How it came about
Nvidia developed the Grace line to address growing power constraints in data centres. The company announced details about the Superchip in late 2022 and early 2023. It represents Nvidia's expansion into central processing units to complement its graphics processors. The goal was to create a system optimized for data movement and large-scale computation. Early benchmarks and developer resources appeared in 2023 as the product moved toward availability.
Inside the chip
The Superchip contains 144 Arm Neoverse V2 cores, split across two dies. Each core includes private instruction and data caches. A custom Scalable Coherency Fabric connects the cores, providing high bandwidth for internal communication. The module uses LPDDR5X memory, which offers high speed at lower power than standard DDR memory. Error-correcting code ensures data reliability. The two chips link via NVLink Chip-to-Chip technology, allowing them to act as a unified system with shared memory access.
How it was made
Nvidia designs the chip but does not specify the manufacturing foundry in the provided text. The module integrates server-class LPDDR5X memory directly on the package. This approach reduces the power needed for memory operations compared to discrete memory modules. The entire Superchip module, including memory, operates within a 500-watt power envelope. It supports multiple PCIe Gen5 lanes for connecting to other system components like storage or network cards.
Where it went
This hardware targets hyperscale cloud applications, high-performance computing, and AI training. Workloads include weather prediction models, graph analytics, and large-scale data mining. Companies use it for tasks that require moving vast amounts of data quickly. The high memory bandwidth benefits applications like reservoir simulation and microservices orchestration. It serves as the heart of a two-socket server configuration, replacing traditional x86-based systems in specific high-efficiency scenarios.
Versions and family
The Grace portfolio includes the single-socket Grace C1 for edge and cloud deployments. The C1 variant features 72 cores and lower power consumption options. The Superchip is the dual-socket variant with 144 cores. Memory configurations vary, with options ranging from 120GB to 960GB of LPDDR5X. The Superchip specifically offers up to 960GB of memory, doubling the capacity available in smaller configurations.
Why it matters
The Superchip challenges traditional x86 server dominance by offering better performance per watt. Nvidia claims it doubles performance at the same power level compared to leading x86 CPUs. It enables data centres to either double their output or cut energy usage in half. This efficiency is critical for managing the rising costs and environmental impact of large-scale computing. It supports the Arm software ecosystem in enterprise environments.
Problems and reception
Industry observers noted the high core count and memory bandwidth as significant features. Comparisons were drawn to contemporary AMD and Intel processors. Some analysts highlighted the potential for Arm architecture to gain ground in the server market. The use of LPDDR5X in a server context was seen as a novel approach to balancing cost and performance. Benchmarks showed strong results in specific scientific and analytic tasks.
Questions people ask about Grace CPU Superchip
How many cores does the Grace CPU Superchip have?
The Grace CPU Superchip contains a total of 144 Arm Neoverse V2 cores. These are distributed across two connected CPU chips within a single module.
What type of memory does it use?
It uses server-class LPDDR5X memory. This type provides high bandwidth and lower power consumption compared to traditional DDR5 memory used in many servers.
What is the power consumption of the Superchip?
The entire Superchip module, including the CPUs and memory, consumes up to 500 watts. This is considered efficient for the level of performance and memory bandwidth provided.
Who designed the Grace CPU Superchip?
Nvidia designed the Grace CPU Superchip. It is part of their broader strategy to provide complete solutions for AI and high-performance computing data centres.
Good reading and sources
There is no English Wikipedia article on Grace CPU Superchip yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/grace-cpu-superchip/, servethehome.com/nvidia-grace-superchip-features-144-cores-and-128-pci, nextplatform.com/compute/2022/03/25/nvidia-embraces-the-cpu-world-with. 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.