Nvidia recently unveiled its DGX GB200 NVL72 supercomputer-in-a-rack at Nvidia GTC 2024 and Patrick Kennedy at Serve The Home took a selection of great photos showcasing the impressive beast.
The name of the DGX GB200 NVL72 tells you much of what you need to know. The GB200 signifies the Grace Blackwell GB200 compute structure, while the NVL72 denotes there are 72 Blackwell GPUs connected by NVLink.
The Blackwell platform contains 208 billion transistors across its two GPU dies. These are connected by 10 TB/second chip-to-chip link into a single, unified GPU. Blackwell, set to ship later this year, will reportedly offer up to 20 petaflops of FP4 power and be up to 30x faster than Hopper for AI inference tasks.
120kW power load
The rack scale system comprises ten compute nodes in the top stack, each featuring dual Infiniband ports, four E1.S drive trays, and management ports. Each node is powered by two Grace Arm CPUs connected to two Blackwell GPUs. Below these nodes are nine NVSwitch shelves, with gold handles for easy removal.
The rear of the rack reveals the power delivery system designed for blind-mate power via the bus bar, liquid cooling nozzles, and NVLink connections for each component. This setup allows for slight movement to ensure proper blind mating.
DGX GB200 NVL72 weighs 1.36 metric tons (3,000 lbs) and consumes a 120kW, a power load that Serve The Home points out, not all data centers will be able to handle. As many can only support a maximum of 60kW racks, a future half-stack system seems a possibility. The rack uses 2 miles (3.2 km) of copper cabling instead of optics to lower the system’s power draw by 20kW.
You can view the rest of the photos taken by Kennedy at GTC 2024 here.