Built for trillion-parameter workloads, the XA GB721-E2 fuses Grace and Blackwell inside a liquid-cooled 48U rack for fast deployment and datacenter sanity.
ASUS is rolling out the XA GB721-E2, a rack-scale platform based on NVIDIA GB300 NVL72 to push model training, high-throughput inference, and HPC past the usual bottlenecks. In one NVLink domain, you get 36 Grace CPUs and 72 Blackwell Ultra GPUs wired for low-latency, high-bandwidth GPU-to-GPU traffic, packaged in a serviceable 48U MGX rack with liquid cooling aimed at predictable performance.
Rack-Scale Muscle, One NVLink Domain
At the heart is a full NVLink fabric spanning 72 Blackwell Ultra GPUs and 36 Grace CPUs. That layout targets trillion-parameter jobs where inter-GPU chatter kills throughput. By collapsing communication into a single domain, the system trims hops, flattens latency, and keeps tensor cores fed under sustained load.
Liquid-Cooled, Modular, and Meant to Be Touched
The 48U MGX rack integrates manifold-based liquid cooling, nine NVLink switch trays, and 18 compute trays. The plumbing is about stability and serviceability: predictable thermals, cleaner maintenance windows, and fewer airflow compromises when you pack dense accelerators into one cabinet.
Balanced I/O and Storage for Real Pipelines
Northbound and east-west traffic rides either the NVIDIA Quantum-X800 InfiniBand stack or Spectrum-X Ethernet with ConnectX-8 SuperNIC, so clusters scale without whiplash. For data, ASUS pairs all-flash hot tiers for AI/HPC with hybrid or unified arrays to handle backup and archival, letting teams tune cost vs. performance without bolting on yet another silo.
From Rack Arrival to Production in Minutes
On the ops side, ASUS Control Center (ACC) and Infrastructure Deployment Center (AIDC) promise near zero-touch onboarding, with orchestration that can bring a full rack online in roughly a half hour. It’s a turnkey path from installation to scheduled jobs, with lifecycle management baked in rather than duct-taped on.
For organizations wrestling with validation, cooling loops, and network topologies, ASUS Professional Services covers design through optimization. Pricing is typically on request, and availability rolls out via enterprise channels; ASUS didn’t publish public list prices for this configuration. Learn more at ASUS, and yes, the company is betting hard that a faster time-to-compute beats chasing raw TFLOPS on paper.
Technical Specifications
| Component | Details |
|---|---|
| Platform | NVIDIA GB300 NVL72 rack-scale system |
| CPU | 36 NVIDIA Grace CPUs in a single NVLink domain |
| GPU | 72 NVIDIA Blackwell Ultra GPUs with NVLink |
| Interconnect | NVIDIA NVLink across compute trays; nine NVLink switch trays |
| Networking | NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet; ConnectX-8 SuperNIC |
| Compute Trays | 18 trays within 48U MGX-compliant rack |
| Cooling | Manifold-based liquid cooling for predictable thermals |
| Management | ASUS Control Center (ACC) and ASUS Infrastructure Deployment Center (AIDC) |
| Deployment | Turnkey rollout; full-rack online target ~30 minutes |
| Storage Options | All-flash hot tier for AI/HPC; hybrid and unified arrays for enterprise data |
| Use Cases | Large-scale model training, high-throughput inference, HPC pipelines |
