Intel Habana Gaudi 2 HL-225B-BOX Retail Channel SKU

The Intel Habana Gaudi 2 HL-225B-BOX is a high-performance AI training and inference accelerator built on Intel's second-generation Gaudi architecture. Manufactured by Habana Labs, an Intel company, the Gaudi 2 processor is fabricated on TSMC's 7nm process node and integrates 24 Tensor Processing Cores (TPCs) alongside a Matrix Multiplication Engine (MME) optimized for the dense compute demands of large-scale deep learning. The card is equipped with 96 GB of HBM2e memory across six stacks, delivering 2.45 TB/s of memory bandwidth to sustain high-throughput training runs on large models without frequent data stalling.

A defining characteristic of the Gaudi 2 platform is its integrated high-speed interconnect fabric. The accelerator includes 24 on-die 100 Gb/s RDMA-capable Ethernet ports — 21 used for scale-out server-to-server communication and 3 for host connectivity — eliminating the need for a separate networking switch in many scale-out topologies and reducing infrastructure complexity in AI clusters. This architecture allows dense multi-node training configurations to be assembled with lower latency and reduced cabling overhead compared to solutions that rely entirely on external NVLink or InfiniBand fabrics.

The HL-225B-BOX is the retail channel boxed SKU of the Gaudi 2, intended for deployment in enterprise servers, AI research infrastructure, and datacenter environments where organizations require a discrete, installable accelerator form factor. It is fully supported by Intel's open-source software stack, including the SynapseAI SDK, and is compatible with widely adopted deep learning frameworks such as PyTorch and TensorFlow via Intel's driver and runtime layers. Target workloads include large language model (LLM) training, computer vision model development, recommendation system training, and high-throughput batch inference at scale.

Ideal for

• Large language model (LLM) pre-training and fine-tuning across multi-node GPU clusters using the integrated 100 GbE scale-out fabric
• High-throughput deep learning inference serving for NLP, computer vision, and multimodal AI models in enterprise datacenter deployments
• Recommendation engine training requiring sustained high memory bandwidth to process sparse embedding tables at scale
• Computer vision model development including image classification, object detection, and segmentation on large annotated datasets
• AI research and experimentation environments where open-source framework compatibility (PyTorch, TensorFlow) and developer accessibility are priorities
• Scalable AI infrastructure build-out for cloud service providers and enterprise IT teams seeking an alternative high-performance accelerator to complement existing compute estates

Technical specifications

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