Quanta Computer Intel Gaudi 3 8-OAM HGX-Style Server QCT-G3

The Quanta Computer Intel Gaudi 3 8-OAM HGX-Style Server (QCT-G3-8OAM) is a purpose-built AI training and inference platform integrating eight Intel Gaudi 3 accelerator modules in an OAM (Open Accelerator Module) form factor within an industry-standard HGX-compatible server chassis. Each Gaudi 3 accelerator is built on a 5nm process node and delivers substantial improvements in compute density and memory bandwidth over its predecessor, featuring 64 Tensor Processor Cores (TPCs) and Matrix Multiplication Engines (MMEs) optimized for mixed-precision deep learning operations including BF16, FP8, and FP32. The eight accelerators are interconnected via Intel's high-bandwidth, low-latency scale-up fabric using 24 ports of 200Gbps Ethernet per OAM, enabling direct all-to-all communication without requiring external switch infrastructure for within-node collective operations.

The QCT-G3-8OAM platform is engineered to address the demands of large-scale generative AI model training, large language model (LLM) fine-tuning, and high-throughput inference serving. Each Gaudi 3 OAM carries 128GB of HBM2e memory, yielding an aggregate of 1TB of HBM capacity across the full eight-accelerator configuration, alongside aggregate memory bandwidth well suited to the memory-bound kernels common in transformer-based workloads. The server integrates dual high-core-count Intel Xeon Scalable processors to handle host-side data preprocessing, orchestration, and I/O, with PCIe Gen 5 connectivity between host CPUs and the accelerator complex.

Designed for enterprise datacenter deployment, the QCT-G3 platform supports standard data center power and cooling infrastructure, with the OAM modules utilizing direct liquid cooling or high-airflow air cooling depending on deployment configuration. The system is compatible with Intel's Gaudi software ecosystem, including the Intel Gaudi Software Suite, Optimum Habana integration for Hugging Face workloads, and support for PyTorch and TensorFlow via SynapseAI. Omnixon Global supplies this platform to enterprise and hyperscale customers across the UAE, GCC, EMEA, and APAC regions.

Ideal for

• Large language model (LLM) pre-training and fine-tuning at scale, supporting multi-billion parameter transformer architectures such as LLaMA, Falcon, and GPT-class models
• High-throughput generative AI inference serving for enterprise applications requiring low-latency responses across concurrent user sessions
• Computer vision and multimodal model training for industrial inspection, medical imaging analysis, and autonomous systems development
• Retrieval-augmented generation (RAG) pipeline acceleration combining dense embedding computation with transformer inference in enterprise AI platforms
• Scientific computing and simulation workloads in HPC environments that benefit from high-bandwidth memory and mixed-precision tensor operations
• Enterprise AI model experimentation and MLOps pipelines requiring reproducible, high-throughput training runs with broad deep learning framework compatibility

Technical specifications

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