XA GB721-E2

The XA GB721-E2 is an enterprise rack-scale AI computing system designed for multi-trillion-parameter model training and massive-throughput inference workloads. Manufactured as an NVIDIA MGX-compliant turnkey infrastructure solution (designed by ASUS), the system implements the NVIDIA GB300 NVL72 architecture. It consolidates 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs into a unified 48U rack footprint, acting computationally as a single massive accelerator over a liquid-cooled fabric.

Positioned at the absolute peak of data center AI compute, the XA GB721-E2 moves beyond traditional multi-node server clusters connected over standard InfiniBand fabrics. By tying all 72 GPUs together in a single, hardware-switched NVLink domain, the platform removes standard scale-out communication bottlenecks. It directly targets frontier model pre-training, complex post-training reinforcement learning runs, and high-concurrency inference for agentic pipelines where inter-GPU latency dictates total system throughput.

For engineering teams evaluating the XA GB721-E2 for AI deployments, this rack represents an end-to-end compute factory. With centralized power delivery through a 1400A busbar and dedicated manifold liquid cooling, it is engineered for continuous maximum compute density in modern high-density data centers. It competes directly with hyperscale custom clusters and preceding-generation GB200 NVL72 deployments, delivering significantly expanded memory footprints and superior floating-point throughput.

The defining architectural feature of the XA GB721-E2 is its shared memory architecture and bandwidth density. Across the entire 48U rack, the platform integrates 72 Blackwell Ultra GPUs offering a total of 20,736 GB (20.7 TB) of coherent HBM3e VRAM. The system is split across 18 compute trays, with each tray hosting two "Bianca" compute boards.

• Compute Nodes: 18 compute trays housing 36 NVIDIA Grace CPUs (Neoverse V2 Arm architecture) and 72 NVIDIA Blackwell Ultra GPUs.
• Unified GPU Memory: 20,736 GB of HBM3e VRAM across the rack, with each compute tray providing 576 GB of HBM3e per Bianca board at bandwidths reaching up to 16 TB/s.
• Host System Memory: 480 GB of LPDDR5X ECC memory per Bianca board (up to 512 GB/s bandwidth), providing high-speed system caching and dataset staging.
• Interconnect Fabric: 9 NVLink switch trays establishing a single NVLink domain with 14.4 TB/s aggregate bidirectional switching capacity. Each GPU communicates across the fifth-generation NVLink fabric at 1,800 GB/s.
• Networking and I/O: 4 x NVIDIA ConnectX-8 800GbE OSFP network adapters and 1 x NVIDIA BlueField-3 B3240 DPU per tray, alongside 8 x E1.S NVMe hot-swap storage bays.
• Power and Physical Infrastructure: 48U NVIDIA MGX rack enclosure fed by a single 1400A busbar, backed by either six or eight 33 kW power shelves (supporting over 198 kW to 264 kW capacity), and dual-sided liquid cooling manifolds.

Memory bandwidth is the primary operational constraint in auto-regressive token generation. With individual compute boards transferring data at up to 16 TB/s over HBM3e, the XA GB721-E2 AI inference performance minimizes memory-bound stalls during the decode phase. Furthermore, because all 72 GPUs exist within a non-blocking NVLink domain, tensor parallelism and pipeline parallelism scale across all 72 chips with negligible latency penalty compared to traditional PCIe or RDMA network fabrics.

With a 20736GB GPU for AI workloads, the platform is not restricted by standard model-partitioning bottlenecks. It runs the largest open-weights and proprietary foundation architectures in full precision unquantized FP16 or BF16 across its native NVLink domain.

• DeepSeek-R1 and DeepSeek-V3 (671B parameters): Runs fully in memory at native BF16/FP16 without requiring offloading or heavy quantization. The full 671 billion parameter Mixture-of-Experts (MoE) footprint (requiring approximately 1.3 TB to 1.4 TB of VRAM) fits comfortably, leaving over 19 TB of free capacity for massive key-value (KV) caches, supporting concurrent contexts for thousands of simultaneous streams.
• Llama 3.1 405B: Runs unquantized at native FP16/BF16 (requiring approximately 810 GB of base weight capacity). Tensor parallelism across 8, 16, or 72 GPUs eliminates compute bottlenecks, allowing real-time generation speeds exceeding 100+ tokens per second per stream even under heavy context loads.
• Dense 70B Models (Llama 3.3 70B, Qwen 2.5 72B): A single 70B model requires only ~140 GB in BF16. The XA GB721-E2 can host dozens of full 70B instances concurrently, serving independent enterprise workloads or large-scale synthetic data generation pipelines across isolated GPU slices.

While standard multi-GPU nodes rely on 4-bit (INT4) or 8-bit (FP8/INT8) quantization to run 405B or 671B parameter models, the XA GB721-E2 VRAM for large language models allows operators to run FP8 or FP4 purely for raw throughput gains rather than out of memory necessity.

Using native Blackwell FP4 and FP8 Tensor Cores, the XA GB721-E2 tokens per second scale exponentially on high-throughput serving engines like TensorRT-LLM and vLLM. The massive unified VRAM allows running 1M+ token context windows natively on architectures such as Qwen 2.5 and Llama 3.1, maintaining large KV-caches in high-bandwidth HBM3e without offloading to slower CPU memory. Multimodal foundation models processing concurrent high-resolution video streams and high-dimensional embeddings run concurrently without context truncation.

The XA GB721-E2 is not built for hobbyists running local chatbots or single workstations; it is a data-center-scale compute system. It targets specific engineering and enterprise requirements:

• Enterprise AI Infrastructure Teams: Organizations setting up private AI factories to maintain sovereignty over proprietary data, training internal LLMs, and serving internal model APIs.
• Autonomous Multi-Agent Orchestration: Teams building agentic workflows where hundreds of interconnected agents continuously query large 70B and 405B models. The near-zero latency of the unified NVLink domain eliminates the latency spikes common in network-distributed agent systems.
• AI Research Labs and Model Developers: Practitioners conducting full pre-training runs on domain-specific architectures or executing compute-intensive post-training methods (such as PPO, GRPO, and direct preference optimization) that demand continuous gradient synchronization across many nodes.
• High-Throughput API Providers: Organizations needing maximum token throughput per square foot of data center space, where running models in FP8/FP4 at massive batch sizes translates directly into reduced cost-per-token.

Evaluating the XA GB721-E2 requires comparing it against both its direct predecessor architectures and alternative multi-node HGX deployments.

The GB300 NVL72 platform updates the earlier GB200 NVL72 architecture primarily through memory subsystem improvements and higher compute density. While the GB200 NVL72 provided 13,824 GB of total HBM3e across 72 GPUs, the XA GB721-E2 steps up to 288 GB HBM3e per GPU (576 GB per Bianca board), providing 20,736 GB in the rack. This 50 percent increase in VRAM capacity allows model serving deployments to retain massive KV caches at full precision without sharding across multiple racks.

Achieving 72 GPUs using standard 8-way HGX server nodes requires 9 separate chassis linked over InfiniBand or RoCE switches.

• Interconnect Overhead: Traditional HGX clusters rely on network transceivers and leaf-spine switches for cross-chassis communication, introducing network latency and routing overhead during All-to-All communication phases. The XA GB721-E2 handles all 72 GPUs over copper cartridge NVLink switches within the rack at 1.8 TB/s per GPU, providing significantly lower latency.
• Footprint and Power: 9 discrete air-cooled or hybrid nodes consume significantly more rack units, external cabling, and cooling infrastructure. The XA GB721-E2 integrates compute, networking, and power into a single 48U liquid-cooled footprint, lowering total operational complexity for sites prepared for 100 kW+ per-rack liquid cooling infrastructure.

For teams deploying hardware for running 70B parameter models at enterprise scale, or serving 405B to 671B foundation systems without distributed networking bottlenecks, the XA GB721-E2 provides unmatched single-rack density and bandwidth.