Gigabyte NVIDIA GB200 NVL2 Compute Tray

The Gigabyte NVIDIA GB200 NVL2 Compute Tray is a high-density, rack-scale AI compute solution built on NVIDIA's Blackwell architecture. The NVL2 form factor integrates two GB200 GPUs and one Grace CPU into a single NVLink-connected tray, enabling a unified, high-bandwidth memory and compute fabric across the combined unit. With NVLink interconnect bandwidth of 1.8 TB/s between the two GPUs, the system eliminates traditional PCIe bottlenecks for large-scale model parallelism and memory-intensive AI workloads.

Each GB200 GPU within the NVL2 tray delivers up to 1,000 TFLOPS of FP8 AI compute performance and incorporates 192 GB of HBM3e memory per GPU, yielding 384 GB of aggregate GPU memory across the tray. The Grace CPU component provides 72 Arm Neoverse V2 cores and 480 GB of LPDDR5X system memory, connected to the GPUs via NVIDIA's NVLink-C2C chip-to-chip interconnect at 900 GB/s of total bidirectional bandwidth. This unified architecture allows models to span CPU and GPU memory transparently, supporting inference and training of frontier-scale large language models that exceed the capacity of conventional GPU servers.

Designed for integration into NVIDIA GB200 NVL72 or compatible liquid-cooled rack infrastructure, the Gigabyte GB200 NVL2 Compute Tray is engineered for direct liquid cooling to manage the extreme thermal output of Blackwell-generation silicon. It is intended for enterprise data centers, cloud service providers, and AI research institutions requiring maximum compute density per rack unit for generative AI training, large-scale inference, and scientific simulation workloads.

Ideal for

• Training and fine-tuning frontier-scale large language models (LLMs) with hundreds of billions to trillions of parameters requiring distributed tensor and pipeline parallelism
• High-throughput generative AI inference serving for enterprise-deployed foundation models, minimizing latency through massive on-chip memory capacity
• Scientific high-performance computing (HPC) simulations in computational fluid dynamics, molecular dynamics, and climate modeling that demand extreme floating-point throughput
• Retrieval-augmented generation (RAG) pipelines and vector database acceleration requiring fast, large-memory GPU execution
• Multi-modal AI model development including vision-language and video generation models that require simultaneous high memory bandwidth and compute throughput
• Enterprise AI infrastructure consolidation within liquid-cooled hyperscale data centers seeking maximum AI compute density per rack

Technical specifications

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