Scaling Everpure NVIDIA Enterprise Reference Architecture with NVIDIA OVX and HGX Servers for Different Deployments

PCIe Optimized 2-4-3-200 Deployment

The PCIe Optimized 2-4-3-200 reference configuration is an Enterprise RA design pattern for scale-out compute nodes, featuring:

• 2x CPUs for balanced processing

• 4x PCIe GPUs (eligible GPUs:NVIDIA RTX PRO 6000 Blackwell Server Edition, H100 NVL, H200 NVL and L40S) for accelerated compute

• 3x network adapters/DPUs (e.g., BlueField-3) for high-speed connectivity

• 200Gbps network bandwidth per GPU

Use cases include:

• AI inference: Medium model parameter inference workloads

• AI training: Small model training and fine-tuning

PCIe Optimized 2-8-5-200 Deployment

The PCIe optimized 2-8-5-200 reference configuration is a scale-out compute node design, optimized for AI and high-performance workloads. Key features include:

• 2x CPUs for balanced processing

• 8x PCIe GPUs (eligible GPUs: NVIDIA RTX PRO Blackwell 6000 Server Edition, H100 NVL, H200 NVL and L40S) for accelerated compute

• 5x high-speed network adapters/DPUs (e.g., BlueField-3)

• 200Gbps network bandwidth per GPU

HGX 2-8-9-400 Reference Configuration

The HGX Reference Configuration 2-8-9-400 is a design pattern for high-performance AI and HPC workloads, featuring:

• 2x CPUs (e.g., Intel Xeon 8480C PCIe Gen5 or AMD Turin) for balanced processing

• 8x GPUs (eligible GPUs: HGX H100/H200/B200) with 4th-gen NVLink, delivering 900 GB/s GPU-to-GPU bandwidth

• 9x network adapters (e.g., BlueField-3) supporting up to 400Gbps per GPU

Use cases include:

• AI inference: Large (per node) and medium (per GPU) model parameter inference workloads

• AI training: Large to small model training and fine-tuning based on cluster sizing