Supermicro AS-2125GS-TNHR GPU SuperServer (Dual AMD EPYC, 4x H100 NVL)

The Supermicro AS-2125GS-TNHR is a 2U dual-socket AMD EPYC GPU SuperServer engineered to deliver exceptional parallel compute density for artificial intelligence inference, training, and high-performance computing workloads. Built on AMD's EPYC 9004 series (Genoa) processor platform, the system accommodates two EPYC CPUs supporting PCIe 5.0 interconnects and offers high memory bandwidth to feed demanding GPU workloads. Four NVIDIA H100 NVL GPUs are housed within the 2U chassis, providing access to NVIDIA's Hopper architecture and NVLink fabric for high-bandwidth GPU-to-GPU communication, making this system well suited for large language model (LLM) inference and distributed AI compute tasks.

The server is designed around Supermicro's server management ecosystem and supports out-of-band management via IPMI 2.0 with a dedicated BMC. Connectivity is addressed through onboard 10GbE LAN ports alongside high-speed network expansion options, while multiple PCIe 5.0 expansion slots accommodate additional networking or storage accelerators. The system supports DDR5 registered ECC memory across multiple DIMM slots per processor, enabling large in-memory datasets critical to AI inference pipelines and HPC simulation tasks. Storage flexibility is provided through NVMe and SATA/SAS drive bays, accommodating fast local storage for model weights and datasets.

The AS-2125GS-TNHR has gained traction in enterprise and hyperscale AI infrastructure deployments as a computationally dense, space-efficient building block for GPU clusters. Its balance of dual EPYC CPU compute, four H100 NVL accelerators, DDR5 memory capacity, and PCIe 5.0 I/O bandwidth makes it a practical choice for organizations constructing on-premises AI inference infrastructure, scientific computing clusters, and data analytics platforms where GPU throughput and CPU-GPU data transfer performance are primary considerations.

Ideal for

• Large language model (LLM) inference serving, where the four H100 NVL GPUs and NVLink interconnect reduce token generation latency at scale
• Distributed AI model training across multi-node GPU clusters, leveraging PCIe 5.0 and high-bandwidth networking fabrics
• High-performance computing simulations in scientific research, engineering, and computational fluid dynamics requiring dense parallel processing
• AI-driven data analytics and real-time inference pipelines processing high-throughput data streams within enterprise data centers
• Generative AI application hosting, including text, image, and multimodal model serving requiring large GPU memory footprints provided by H100 NVL
• On-premises AI infrastructure build-outs where organizations consolidate GPU compute into space-efficient 2U nodes to maximize rack density

Technical specifications

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