Inventec P-G2-8OAM Gaudi 2 Server Platform

The Inventec P-G2-8OAM is a purpose-built server platform integrating eight Intel Gaudi 2 AI accelerators in OAM (Open Accelerator Module) form factor, designed to address the demanding computational requirements of large-scale deep learning training and high-throughput AI inference workloads. Built around Intel's Gaudi 2 architecture, the platform delivers substantial matrix multiplication throughput via dedicated Tensor Processing Cores (TPC) and Matrix Multiplication Engines (MME), enabling enterprises to run foundation model training, generative AI, and complex neural network workloads at scale. The system conforms to OAM consortium specifications, facilitating interoperability within open accelerator ecosystems.

Each of the eight Gaudi 2 OAM modules integrated into the P-G2-8OAM incorporates 96 GB of HBM2E memory per accelerator, providing the high-bandwidth memory capacity essential for fitting large model parameters and activations in-device. The accelerators are interconnected via Intel's high-speed HCCL (Habana Collective Communications Library)-compatible fabric using 21 integrated 100 Gbps RDMA-capable Ethernet ports per Gaudi 2 device, enabling efficient all-reduce and collective communication operations without dependency on external networking ASICs for scale-out. This architecture supports both scale-up within the node and scale-out across multi-node clusters.

The Inventec P-G2-8OAM platform is engineered for enterprise datacenter deployment, supporting standard datacenter power and cooling infrastructure. It is compatible with Intel's Gaudi software ecosystem including the Habana SynapseAI SDK, enabling integration with leading AI frameworks such as PyTorch and TensorFlow. The platform is positioned for organizations building private AI infrastructure, sovereign AI deployments, and high-performance computing environments across datacenter and cloud-adjacent facilities in regions including UAE, GCC, EMEA, and APAC.

Ideal for

• Large language model (LLM) pre-training and fine-tuning for enterprise generative AI initiatives requiring high-memory-bandwidth multi-accelerator nodes
• High-throughput AI inference serving for production deployment of transformer-based models including text generation and multimodal workloads
• Computer vision model training at scale, including object detection, segmentation, and classification pipelines for industrial and retail applications
• Recommender system training for large embedding table workloads in e-commerce, financial services, and digital media platforms
• Sovereign and private AI datacenter buildout for government and regulated-industry organizations requiring on-premises control of AI compute infrastructure
• Multi-node distributed deep learning research and development clusters leveraging integrated scale-out networking to eliminate external switch dependencies

Technical specifications

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