Intel Gaudi 2 HL-225B PCIe Accelerator Card 96GB HBM2e

The Intel Gaudi 2 HL-225B is a PCIe Gen 4.0 accelerator card built on Intel's second-generation Gaudi architecture, designed to deliver high-throughput deep learning training and inference performance for enterprise datacenters. Featuring 96GB of HBM2e memory across six stacks with 2.45 TB/s of aggregate memory bandwidth, the HL-225B provides the memory capacity and bandwidth required to accommodate large language models, computer vision networks, and recommendation systems without model parallelism overhead. The card communicates over a standard PCIe x16 Gen 4.0 host interface, enabling broad server compatibility across major platforms without requiring proprietary interconnect infrastructure.

The Gaudi 2 architecture integrates 24 fully programmable Tensor Processor Cores (TPCs) alongside a Matrix Multiplication Engine (MME) optimized for GEMM operations central to transformer-based model training. Eight on-die 100 Gigabit Ethernet (RoCE v2) RDMA ports enable scale-out networking directly from the silicon, allowing multi-node training clusters to be constructed using standard Ethernet switching infrastructure rather than proprietary fabrics. The HL-225B variant routes these ports externally through an OSFP connector, supporting direct integration into high-bandwidth datacenter network topologies.

The card is supported by Intel's SynapseAI software suite, which provides a PyTorch and TensorFlow integration layer, a graph compiler, and runtime libraries that allow enterprises to migrate existing AI workloads with minimal code changes. The HL-225B is positioned for organizations seeking an alternative to dominant GPU incumbents for large-scale AI training infrastructure, particularly where open networking standards and total infrastructure flexibility are operational priorities. Omnixon Global supplies the HL-225B to enterprise buyers across the UAE, GCC, EMEA, and APAC regions.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale, leveraging 96GB HBM2e capacity to fit multi-billion parameter models on fewer accelerator cards
• Enterprise AI inference serving for transformer-based models including BERT, GPT variants, and vision transformers requiring high memory bandwidth and low latency
• Recommendation engine training for e-commerce, media, and financial services platforms where embedding table size demands substantial on-card HBM capacity
• Computer vision model training for manufacturing quality inspection, medical imaging analysis, and autonomous systems using large batch sizes enabled by ample memory headroom
• Multi-node distributed training clusters built over standard 100GbE RoCE v2 Ethernet fabric, eliminating dependency on proprietary high-speed interconnects
• Datacenter AI infrastructure modernization projects requiring PCIe form-factor accelerators compatible with existing server chassis and power delivery systems

Technical specifications

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