Intel Gaudi 3 HL-325H OAM Accelerator 128GB HBM2e

The Intel Gaudi 3 HL-325H is an OAM (OCP Accelerator Module) form-factor AI accelerator built on Intel's third-generation Gaudi architecture, fabricated on a 5nm process node. The HL-325H integrates 128GB of HBM2e memory across eight HBM2e stacks, delivering 3.7TB/s of aggregate HBM memory bandwidth. The accelerator features 64 Tensor Processor Cores (TPCs) and a dedicated Matrix Multiplication Engine (MME), providing substantial throughput for both training and inference workloads across FP8, BF16, FP16, and FP32 precisions. With 1835 TOPS of FP8 peak compute, the HL-325H is positioned for demanding large-scale AI model development in high-density datacenter deployments.

The Gaudi 3 architecture incorporates 24 integrated 100GbE RoCE v2 network ports directly on-die, enabling scale-out communication between accelerator nodes without requiring third-party networking silicon. This architecture supports direct server-to-server connectivity at up to 2.4Tb/s of total bisectional network bandwidth per accelerator, reducing fabric latency and infrastructure complexity in large AI clusters. The OAM mechanical form factor conforms to OCP OAM specifications, facilitating integration into OAM-compliant Universal Baseboard (UBB) systems from multiple platform vendors.

The HL-325H is supported by Intel's Gaudi software ecosystem, including the Habana SynapseAI SDK, which provides integration with industry-standard frameworks including PyTorch and TensorFlow. The accelerator is suited for enterprises deploying foundation model training, large language model fine-tuning, and high-throughput inference serving at scale across datacenter and cloud infrastructure environments in regulated and performance-sensitive industries.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale using multi-node Gaudi 3 clusters interconnected via on-chip RoCE v2 networking
• High-throughput generative AI inference serving for enterprise applications requiring low-latency, high-concurrency response workloads
• Computer vision and multimodal model training for industries including healthcare imaging, autonomous systems, and retail analytics
• Recommendation system training and serving for large-scale e-commerce, media streaming, and personalization platforms
• Scientific and research AI workloads in energy, genomics, and climate modeling that require high HBM bandwidth and FP32/BF16 compute precision
• Enterprise MLOps pipelines requiring dense GPU-equivalent compute within OAM-compatible high-density baseboard server chassis

Technical specifications

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