Intel Gaudi 2 8-Card OAM Baseboard HLS-GAUDI2-8-OAM

The Intel Gaudi 2 8-Card OAM Baseboard (HLS-GAUDI2-8-OAM) is a high-density AI training and inference platform built around eight Intel Gaudi 2 accelerator modules in the Open Accelerator Module (OAM) form factor. Each Gaudi 2 die is fabricated on TSMC 7nm process technology and integrates 96 Tensor Processor Cores (TPCs) alongside a Matrix Multiplication Engine (MME), delivering substantial throughput for deep learning matrix operations and general-purpose tensor workloads. The baseboard interconnects all eight OAM modules via a fully non-blocking, on-board 24-port 100GbE RoCE v2 fabric using integrated RDMA NICs, enabling high-bandwidth, low-latency peer-to-peer communication without a discrete networking switch for scale-up configurations.

Each Gaudi 2 OAM module carries 96 GB of HBM2e memory, yielding a total of 768 GB of aggregate HBM2e capacity across the eight-module baseboard. Memory bandwidth per module reaches 2.45 TB/s, providing the memory throughput necessary to sustain large model training runs involving billions of parameters. The platform supports PCIe Gen 4 x16 host connectivity and is designed for integration into standard OCP-compliant server chassis, making it suitable for hyperscale and enterprise datacenter deployments that follow open hardware specifications.

The HLS-GAUDI2-8-OAM targets large-scale generative AI, large language model (LLM) training, computer vision, and recommendation system workloads. Software support is provided through the Intel Gaudi Software Suite, which includes the SynapseAI SDK, a PyTorch integration layer, and support for Hugging Face model libraries, allowing data science teams to migrate existing GPU-based training pipelines with minimal code modification. The platform is positioned as a datacenter-grade solution for enterprises and cloud service providers in UAE, GCC, EMEA, and APAC regions seeking high-performance AI compute infrastructure.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale, including transformer-based architectures with billions of parameters requiring high HBM bandwidth and fast all-reduce communication
• Generative AI model development for text, image, and multimodal applications where sustained tensor throughput and large on-board memory capacity are critical
• Enterprise computer vision training pipelines, including object detection, segmentation, and image classification models running across distributed multi-node GPU clusters
• Deep learning-based recommendation system training used by e-commerce, media, and fintech enterprises processing large embedding tables that benefit from high aggregate HBM2e capacity
• AI inference serving for large models in production datacenter environments where low-latency, high-throughput batch inference is required at enterprise scale
• Research and development environments in academia, national laboratories, and corporate AI labs requiring an open, PCIe and OCP-compliant accelerator platform compatible with standard HPC server infrastructure

Technical specifications

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