Gigabyte NVIDIA GB200 NVL72 Rack-Scale GPU System

The Gigabyte NVIDIA GB200 NVL72 Rack-Scale GPU System is a full-rack AI supercomputing platform built around NVIDIA's Blackwell architecture, integrating 72 NVIDIA B200 Tensor Core GPUs and 36 NVIDIA Grace CPUs within a single NVLink-domain chassis. The system leverages fifth-generation NVLink to connect all 72 GPUs at an aggregate 130 TB/s of GPU-to-GPU bandwidth, enabling the entire rack to function as a single, unified compute entity rather than a collection of discrete nodes. With a total of 1.4 ExaFLOPS of AI performance (FP4) and 13.5 TB of unified HBM3e memory across the rack, the GB200 NVL72 is purpose-engineered for the most demanding large-scale model training and inference workloads in production datacenters.

At the hardware level, each B200 GPU in the system delivers up to 20 PetaFLOPS of FP4 Tensor Core performance and 192 GB of HBM3e memory operating at 8 TB/s of memory bandwidth per GPU. The Grace CPUs provide high-bandwidth, low-latency coherent memory interconnects via NVLink-C2C, eliminating traditional PCIe bottlenecks between CPU and GPU domains. The rack integrates a liquid-cooled thermal design capable of supporting the extreme power densities inherent to this class of system, with per-rack TDP in the range defined by NVIDIA's NVL72 reference platform. Networking is served by NVIDIA Quantum-X800 InfiniBand or Spectrum-X800 Ethernet fabric options for multi-rack scale-out.

The GB200 NVL72 targets hyperscale cloud operators, national AI research institutions, large financial services firms, and enterprise customers building sovereign AI infrastructure. It is optimized for trillion-parameter foundation model training, real-time large language model inference at scale, physics-based simulation, and high-throughput data analytics pipelines. Gigabyte's implementation adheres to the NVIDIA MGX reference architecture, ensuring ecosystem compatibility with NVIDIA AI Enterprise software, CUDA libraries, and the broader NVIDIA accelerated computing stack.

Ideal for

• Large-scale foundation model training (LLM, multimodal, and diffusion models with trillions of parameters) requiring unified high-bandwidth GPU memory across the full rack
• High-throughput generative AI inference serving, enabling real-time responses for enterprise-grade AI applications with minimal latency at massive concurrency
• Sovereign AI datacenter buildout for government and national research bodies requiring on-premises, rack-scale accelerated compute capacity
• Scientific high-performance computing including climate modeling, molecular dynamics simulation, and computational fluid dynamics at previously infeasible scales
• Financial services risk modeling and quantitative analytics workloads demanding extreme floating-point throughput and large in-memory dataset processing
• Healthcare and life sciences genomics, drug discovery, and medical imaging AI pipelines that benefit from unified memory capacity and NVLink-accelerated data movement

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.