Skip to main content

Features

Performance, operational efficiency, and fabric reliability were the primary considerations in the design of the rack solution.

The rack solution implements the following features and technologies:

  • High HBM3e bandwidth

    The rack solution utilizes HBM3e technology to deliver a total of 20.7 TB of HBM3e across the 72-GPU NVLink domain. This allows the system to process large AI models instantly and handle massive scientific data.

  • High network bandwidth

    The rack solution features high-speed networking via NVIDIA BlueField-3 DPU adapters or ConnectX-8 SuperNICs. This integrated architecture provides massive aggregate bandwidth across the entire rack, ensuring seamless data flow for large-scale AI workloads.

  • Large system-memory capacity

    The system utilizes a unified memory architecture featuring LPDDR5X memory with error correcting code (ECC) integrated directly onto the NVIDIA Grace processor. This provides up to 17.3 TB of system memory per rack.

  • Large data-storage capacity and hot-swap capability

    Each compute tray supports high-speed E1.S drives, providing localized data staging. With the hot-swap feature, you can add, remove, or replace solid-state drives without turning off the compute tray.

  • Lightpath Diagnostics

    Lightpath Diagnostics provides LEDs to help you diagnose problems. For more information about the Lightpath Diagnostics, see Troubleshooting by system LEDs and diagnostics display.