What Is GB200 Superchip?_
The GB200 Superchip integrates a Grace CPU and two Blackwell GPUs via a high-speed NVLink-C2C interconnect, enabling unified memory access and reduced data movement latency. It is designed for dense AI training and inference workloads, replacing traditional PCIe-based GPU-CPU connections. Each superchip operates as a single logical node within the larger NVL72 rack configuration.
Technical Details
The GB200 Superchip uses NVLink-C2C to achieve chip-to-chip bandwidth that exceeds PCIe Gen5, with the CPU and GPUs sharing a coherent memory pool. In the NVL72 rack, 36 superchips are interconnected via NVLink copper backplane to form a single 72-GPU domain. The superchip's thermal design requires liquid cooling for sustained performance, with cold plates directly mounted on the Grace CPU and Blackwell GPUs. Power delivery is handled through dedicated high-current bus bars per superchip tray, per the OEM specification.
How Leviathan Systems Works with GB200 Superchip
Leviathan field crews install GB200 superchip trays into the NVL72 rack, connecting each tray's NVLink backplane connectors and liquid cooling quick-disconnects. We verify the superchip's power and network cabling per the rack integration guide, ensuring no torque is applied to push-pull connectors.
Related service
GPU Rack Assembly & Rack-and-Stack →