NVIDIA HGX H100 8-GPU SXM5 Baseboard Module

The NVIDIA HGX H100 8-GPU SXM5 Baseboard Module (920-23686-2530-001) is a high-density compute substrate designed to deliver maximum AI and HPC throughput in enterprise data center environments. Built around eight NVIDIA H100 SXM5 GPUs interconnected via fourth-generation NVLink, the baseboard provides a unified, tightly coupled compute fabric that dramatically reduces inter-GPU communication latency and maximizes bandwidth for large-scale parallel workloads. Each H100 SXM5 GPU is manufactured on TSMC's 4N process node and integrates the Hopper architecture, featuring the Transformer Engine with FP8 precision support, second-generation Multi-Instance GPU (MIG) technology, and confidential computing capabilities at the hardware level.

The eight GPUs on the baseboard are connected through NVLink 4.0, delivering 900 GB/s bidirectional bandwidth per GPU and a total aggregate NVLink bandwidth of 7.2 TB/s across the full baseboard. This level of interconnect performance makes the HGX H100 SXM5 baseboard the foundational building block for large language model (LLM) training, multi-node AI inference clusters, and memory-intensive scientific computing tasks that cannot be efficiently partitioned across discrete, loosely coupled accelerators. The module is designed for integration into qualified server platforms and pairs with NVSwitch-based fabric for multi-node scale-out via NVLink Switch Systems.

From a precision computing standpoint, each H100 SXM5 GPU delivers up to 3,958 TFLOPS of FP8 sparse tensor performance, 1,979 TFLOPS of FP8 dense, 989 TFLOPS of FP16 dense Tensor Core operations, and 67 TFLOPS of FP64 Tensor Core performance — specifications that position this baseboard module as a purpose-built solution for the most demanding generative AI training runs, reinforcement learning workloads, molecular dynamics simulations, and real-time large-scale inference deployments across UAE, GCC, EMEA, and APAC data center infrastructure.

Ideal for

• Large language model (LLM) and foundation model training requiring tightly coupled multi-GPU memory and compute fabric across hundreds of billions of parameters
• High-throughput generative AI inference serving for enterprise-scale deployments of models such as GPT-class, diffusion, and multimodal architectures
• High-performance computing (HPC) simulations including computational fluid dynamics, climate modeling, and molecular dynamics that demand FP64 double-precision throughput
• Data center consolidation using NVIDIA Multi-Instance GPU (MIG) to partition each H100 into up to seven isolated GPU instances, enabling secure multi-tenant AI workloads
• Confidential AI computing scenarios where hardware-level TEE (Trusted Execution Environment) isolation is required for regulated industries such as financial services and healthcare
• Large-scale recommendation system training and real-time ranking inference where aggregate GPU memory capacity and NVLink bandwidth eliminate CPU-GPU data transfer bottlenecks

Technical specifications

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