This is what Nvidia’s Exaflop supercomputer-in-a-rack looks like: The DGX GB200 NVL72 tower most likely uses 48V, 2.5kA to deliver a whopping 1,440 petaflops, which could cost millions
Nvidia recently unveiled its DGX GB200 NVL72 supercomputer-in-a-rack at Nvidia GTC 2024 and Patrick Kennedy at Serve it house has captured a selection of amazing photos showing the impressive beast.
The name of the DGX GB200 NVL72 tells you a lot of what you need to know. The GB200 represents the Grace Blackwell GB200 compute fabric, while the NVL72 represents 72 Blackwell GPUs connected via NVLink.
The Blackwell platform contains 208 billion transistors spread over the two GPU chips. These are connected by a 10TB/second chip-to-chip link to a single, unified GPU. Coming to market later this year, Blackwell will reportedly offer up to 20 petaflops of FP4 power and be up to 30x faster than Hopper for AI inference tasks.
Power load of 120 kW
The rack-scale system consists of ten compute nodes in the top stack, each with 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 through the busbar, liquid cooling nozzles and NVLink connections for each component. This arrangement allows for slight movement to ensure good blind mating.
DGX GB200 NVL72 weighs 1.36 tons (3,000 lbs) and consumes 120 kW, a power load that Serve it house points out that not all data centers can handle this. Since many can only support racks of up to 60 kW, a future half-stack system seems like a possibility. The rack uses 2 miles (3.2 km) of copper cabling instead of optics to reduce system power consumption by 20 kW.
You can view the rest of the photos taken by Kennedy at GTC 2024 here.