Intel Gaudi 3 8-Card OAM Baseboard HLS-GAUDI3-8-OAM

The Intel Gaudi 3 8-Card OAM Baseboard (HLS-GAUDI3-8-OAM) is a high-density AI training and inference platform built around eight Intel Gaudi 3 accelerator modules in the Open Accelerator Module (OAM) form factor. Each Gaudi 3 die is manufactured on TSMC's 5nm process node and integrates 64 Tensor Processor Cores (TPCs) alongside a Matrix Multiplication Engine (MME), delivering substantial throughput for large-scale deep learning workloads. The baseboard interconnects all eight accelerators via Intel's high-bandwidth on-board fabric, enabling tightly coupled multi-accelerator communication without external switching overhead.

Each Gaudi 3 OAM module is equipped with 96 GB of HBM2e memory, giving the full 8-card baseboard an aggregate of 768 GB of HBM2e across the system. The platform provides 24 integrated 200 GbE RDMA-capable network ports per accelerator (shared across the baseboard), supporting RoCE v2 for scale-out cluster connectivity. This native Ethernet-based scale-out approach eliminates the need for proprietary interconnect hardware, making it straightforward to deploy in standard datacenter network fabrics for distributed training jobs spanning hundreds or thousands of accelerators.

The HLS-GAUDI3-8-OAM is designed for enterprise datacenters and hyperscale AI infrastructure teams running generative AI model training, large language model (LLM) fine-tuning, and high-throughput inference serving. It is supported by the Intel Gaudi Software Suite, which includes the SynapseAI SDK, integration with PyTorch and TensorFlow via Habana plugins, and compatibility with Hugging Face Optimum Habana—enabling teams to migrate existing GPU-based workflows with minimal code changes. The platform suits environments prioritizing total rack efficiency, open standards networking, and software ecosystem flexibility.

Ideal for

• Large language model (LLM) pre-training and fine-tuning across transformer-based architectures such as LLaMA, GPT, and BERT variants at scale
• Generative AI inference serving for enterprise applications requiring high-throughput, low-latency responses from multi-billion parameter models
• Computer vision and multimodal model training for applications in medical imaging, autonomous systems, and industrial inspection
• Distributed deep learning training clusters leveraging the native 200 GbE RoCE v2 scale-out fabric to span multiple baseboards without proprietary networking hardware
• Enterprise MLOps pipelines requiring a PyTorch- and TensorFlow-compatible accelerator platform with SynapseAI SDK integration
• High-performance AI inferencing in sovereign cloud and regulated datacenter environments across UAE, GCC, EMEA, and APAC regions requiring on-premises AI compute

Technical specifications

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