GB200 Grace Blackwell Superchip
Nvidia AI superchip linking Grace CPUs and Blackwell GPUs
The GB200 Grace Blackwell Superchip is a system-on-chip module from Nvidia that combines Grace central processing units with Blackwell graphics processors. It uses NVLink-C2C interconnects to enable high-speed communication for artificial intelligence training and inference in data centres.
- Combines two Blackwell GPUs with one Grace CPU
- Uses NVLink-C2C for high-speed chip-to-chip links
- Supports up to 480 GB LPDDR5X memory
- Enables FP4 AI precision via Transformer Engine
- Forms base of GB200 NVL72 rack-scale system
What it is
The GB200 Grace Blackwell Superchip is a combined processor module designed by Nvidia for accelerated computing. It integrates two Blackwell Tensor Core graphics processing units with one Grace central processing unit. These components connect via the NVLink-C2C interconnect, allowing them to function as a single cohesive unit. This design targets large-scale artificial intelligence workloads, including real-time inference for trillion-parameter language models. The superchip serves as the fundamental building block for larger rack-scale systems like the GB200 NVL72.
Inside the chip
The architecture merges Arm-based Grace CPUs with Blackwell GPUs. A key feature is the NVLink-C2C interface, which provides high-bandwidth, low-latency communication between the processor types. The Blackwell component includes a second-generation Transformer Engine that supports FP4 and FP8 precision formats. This engine optimises throughput for AI inference and training tasks. The design prioritises energy efficiency and memory bandwidth, aiming to reduce communication bottlenecks often found in distributed GPU clusters. The superchip supports up to 480 GB of LPDDR5X memory with bandwidth reaching 512 GB/s.
Where it went
This superchip powers data centre infrastructure for artificial intelligence and high-performance computing. It is used in systems designed for training large language models and running real-time inference. The GB200 NVL72 rack system employs thirty-six Grace CPUs and seventy-two Blackwell GPUs, built from these superchips. It handles mixture-of-experts architectures and database queries with compression. The technology also supports scientific computing applications through the GB200 NVL4 variant, which links four GPUs with two CPUs.
Versions and family
The GB200 NVL4 is a variant that connects four Blackwell GPUs with two Grace CPUs. It is compatible with liquid-cooled modular servers and targets scientific computing and AI training. The GB300 NVL72 is a related platform featuring Blackwell Ultra GPUs and Grace CPUs, designed for test-time scaling inference. These variants extend the core superchip design into different rack configurations and performance tiers for specific enterprise needs.
Questions people ask about GB200 Grace Blackwell Superchip
What components make up the GB200 Superchip?
The GB200 Superchip contains two Nvidia Blackwell Tensor Core GPUs and one Nvidia Grace CPU. These parts are linked together using the NVLink-C2C interconnect to act as a unified processing unit for AI workloads.
What type of memory does the GB200 use?
The superchip supports up to 480 GB of LPDDR5X memory. This memory configuration provides a bandwidth of up to 512 GB/s, facilitating high-throughput data access for large language models and complex computations.
How does the GB200 connect to other chips?
It uses the NVLink-C2C interconnect to link the Grace CPU and Blackwell GPUs within the superchip. For larger systems, it utilises fifth-generation NVLink and NVLink Switch Systems to connect multiple GPUs across a rack.
Good reading and sources
There is no English Wikipedia article on GB200 Grace Blackwell Superchip yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/gb200-nvl72/, tomshardware.com/pc-components/gpus/nvidia-introduces-a-new-merged-cpu. 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.