Squadron
Complete liquid-cooled rack, 18 compute trays
Allied Squadron NVL72
72 GPUs. One system.
NVIDIA GB300 NVL72 rack-scale AI.
SKU ALL-DLB2-CB3
Plan your Squadron deployment
Overview
Squadron NVL72 is a complete, liquid-cooled NVIDIA GB300 NVL72 rack: 72 Blackwell Ultra GPUs and 36 Grace CPUs joined by NVLink into a single accelerator. It is built for the largest AI training and reasoning workloads, and delivered by Allied as a fully integrated, tested rack.
Ideal for
- Sovereign and national AI programmes
- AI factories and GPU cloud providers
- Frontier-scale model training and real-time reasoning inference
What Allied delivers
- Site assessment covering power, liquid cooling, floor loading and network fabric
- Rack-scale integration and burn-in before shipping, then on-site installation and commissioning
- NVIDIA software stack setup, cluster networking (Quantum-X800 InfiniBand or Spectrum-X Ethernet) and ongoing support
- Scales from a single rack to multi-rack pods
Site requirements
- Liquid cooling through an in-row or in-rack coolant distribution unit (CDU). Facility water must meet ASHRAE liquid-cooling water-quality guidelines.
- Power delivered through six 33 kW power shelves to a 54 V DC bus bar
- Floor space for a 600 mm wide, 1,068 mm deep, 2,299 mm tall rack
Specifications
Rack specifications
- Dimensions (L×W×H mm)
- 1,068 x 600 x 2,299
- GPUs
- 72 x NVIDIA Blackwell Ultra GPUs
- CPUs
- 36 x NVIDIA Grace CPUs
- GPU memory
- 20 TB HBM3E, up to 576 TB/s
- Power
- 6 x 1U 33 kW power shelves54 V DC, 1400 A bus bar
- CPU memory
- 17 TB LPDDR5X ECC, up to 14 TB/s
- Compute trays
- 18 x 1U liquid-cooled compute trays
- NVLink switch trays
- 9 x NVIDIA NVLink Switch trays4 x NVLink cable cartridges
- Networking
- NVIDIA ConnectX-8 SuperNICscompatible with NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet
- Management switches
- 2 x out-of-band management switches1 x optional switch for rack-to-rack connection
- Cooling
- Direct liquid coolingin-row or in-rack CDU
Compute tray (each of 18)
- Form factor
- 1U liquid-cooled compute node
- CPU
- 2 x NVIDIA Grace CPUs (72 Arm Neoverse V2 cores each)
- CPU memory
- 480 GB LPDDR5X per CPU (960 GB per tray)
- GPU
- 4 x NVIDIA Blackwell Ultra GPUs
- GPU memory
- 279 GB HBM3E per GPU (1,116 GB per tray)
- Storage
- 8 x E1.S Gen5 NVMe bays1 x M.2 (2280/22110, PCIe Gen5 x4)
- Networking
- 4 x 800 Gb/s OSFP (NVIDIA ConnectX-8 SuperNIC)1 x NVIDIA BlueField-3 DPU1 x 1 Gb/s LAN
- Front I/O
- 1 x USB 3.2 Gen1 (Type-A)1 x Mini-DP1 x RJ451 x MLAN
- Rear I/O
- 4 x NVIDIA NVLink Switch connectors
- Cooling
- 2 x Superchip cold-plate loops8 x 40x40x56mm fans
- OS support
- Ubuntu 22.04.3 (arm64)Red Hat Enterprise Linux 9.3 (aarch64)SUSE Linux Enterprise Server 15 SP5 (aarch64)