Supermicro SYS-821GE-TNHRT 8U 8x Gaudi 3 OAM Server

The Supermicro SYS-821GE-TNHRT is an 8U rackmount server engineered around eight Intel Gaudi 3 OAM (Open Accelerator Module) AI accelerators, purpose-built for large-scale deep learning training and high-throughput AI inference workloads. Each Gaudi 3 OAM module delivers 64 tensor processor cores and 128 GB of HBM2e memory per accelerator, yielding a total of 1 TB of HBM2e across the full eight-accelerator configuration. The platform leverages Intel's second-generation Gaudi 3 architecture, which integrates 24 x 100 GbE RoCE v2 network ports per OAM for direct scale-out fabric connectivity without requiring a separate InfiniBand switch fabric, enabling low-latency, high-bandwidth all-to-all communication between nodes.

The SYS-821GE-TNHRT is built on a dual-socket Intel Xeon Scalable (Sapphire Rapids or successor) processor platform supporting high-speed PCIe 5.0 interconnects between the host CPUs and the Gaudi 3 OAM modules. The chassis accommodates high-capacity DDR5 system memory and multiple NVMe storage bays to sustain the data pipelines demanded by billion-parameter model training. Supermicro's thermal design integrates a high-density, hot-swap fan infrastructure and a rigid airflow path calibrated for sustained full-load operation in standard 8U datacenter rack deployments. Out-of-band management is provided via an IPMI 2.0-compliant Baseboard Management Controller (BMC), supporting IPMI, Redfish, and SMASH CLP interfaces for integration with enterprise datacenter orchestration tools.

This server is validated for the Intel Gaudi software stack, including Intel Gaudi PyTorch integration and the SynapseAI SDK, enabling direct use of standard deep learning frameworks such as PyTorch and TensorFlow without proprietary middleware dependencies. The platform is positioned for organizations building or scaling AI infrastructure across cloud, on-premises datacenter, and hybrid environments, particularly where total cost of ownership, open ecosystem compatibility, and scale-out networking density are primary architectural considerations.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale, leveraging the combined 1 TB HBM2e capacity and high-bandwidth inter-accelerator fabric across eight Gaudi 3 OAM modules
• Distributed deep learning training across multi-node clusters using native 100 GbE RoCE v2 scale-out networking without an external InfiniBand fabric
• High-throughput AI inference serving for enterprise NLP, computer vision, and multimodal models requiring sustained low-latency response at datacenter scale
• Generative AI model development and experimentation, including diffusion models and transformer architectures, using the Intel SynapseAI SDK and PyTorch integration
• HPC and scientific computing workloads that benefit from high-memory-bandwidth accelerators and tightly coupled CPU-accelerator PCIe 5.0 interconnects
• Enterprise AI platform consolidation, replacing multi-chassis GPU configurations with a high-density 8U form factor to reduce rack space and simplify datacenter power and cooling infrastructure

Technical specifications

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