HPE Intel Gaudi 3 OAM 96GB HBM2e Accelerator

The HPE Intel Gaudi 3 OAM 96GB HBM2e Accelerator (P65893-B21) is a high-performance AI training and inference accelerator built on Intel's third-generation Gaudi architecture. Designed for Open Accelerator Module (OAM) form factor deployments, the card integrates 96GB of HBM2e memory across a wide memory bus, delivering the high bandwidth and capacity essential for large-scale deep learning model training, generative AI workloads, and complex inference serving at scale. The Gaudi 3 architecture features an array of Tensor Processor Cores (TPCs) and Matrix Multiplication Engines (MMEs) purpose-built to accelerate transformer-based and convolutional neural network workloads with strong compute efficiency.

Gaudi 3 introduces significant improvements over its predecessor in both compute throughput and interconnect capability. The accelerator supports onboard RoCE v2-based RDMA networking via integrated 200Gb Ethernet ports, enabling direct accelerator-to-accelerator communication across nodes without relying on a separate networking fabric. This tightly coupled networking approach reduces communication bottlenecks during distributed training runs, making it well suited for multi-node clusters scaling to hundreds of accelerators. The OAM form factor enables dense system configurations within HPE's supported server and accelerator tray platforms.

As an HPE-branded solution, P65893-B21 is integrated into HPE's enterprise server and management ecosystem, supporting deployment alongside HPE iLO management, HPE GreenLake cloud services, and HPE's AI-optimized infrastructure portfolio. The accelerator is targeted at organizations building on-premises AI infrastructure for large language model (LLM) training, fine-tuning, and inference, as well as high-performance computing environments where scalable, standards-based AI acceleration is required. Intel's open software ecosystem, including the Intel Gaudi Software suite and SynapseAI framework, provides broad compatibility with PyTorch and other leading ML frameworks.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale across multi-node HPE accelerator clusters
• Generative AI inference serving for enterprise applications requiring high-throughput, low-latency model execution
• Computer vision and multimodal AI model training for healthcare imaging, manufacturing inspection, and media analytics
• High-performance computing workloads combining AI and simulation in research, energy, and scientific computing environments
• Retrieval-augmented generation (RAG) pipeline acceleration for enterprise knowledge management and intelligent search
• Distributed deep learning research and development on on-premises AI infrastructure with open, standards-based tooling

Technical specifications

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