Altus ALT-G2-8P Gaudi 2 8-Accelerator System

The Intel Altus ALT-G2-8P Gaudi 2 8-Accelerator System is a purpose-built, high-density AI training and inference platform integrating eight Intel Gaudi 2 HL-225H mezzanine card accelerators into a single server chassis. Built on Intel's second-generation Gaudi architecture, each Gaudi 2 accelerator features 96 GB of HBM2E memory and a high-bandwidth on-chip GEMM engine optimized for the matrix multiplications that underpin large-scale deep learning. The system leverages an integrated 24-port 100 GbE RoCE v2 fabric across all eight accelerators, enabling low-latency, high-bandwidth all-to-all communication for distributed training without requiring external network switch infrastructure.

The Gaudi 2 architecture is designed around a cluster of Tensor Processing Cores (TPCs) and a Matrix Multiplication Engine (MME), delivering substantial throughput for both FP32 and BF16 precision workloads. Each accelerator exposes dual NUMA-aware PCIe Gen 4 host interfaces and supports the SynapseAI software suite, which provides compatibility with PyTorch and TensorFlow through operator-level graph compilation and optimization. The ALT-G2-8P chassis aggregates 768 GB of total HBM2E capacity across all eight devices, making it suitable for hosting very large model parameter sets in accelerator memory during training runs.

Targeted at enterprise data centers running generative AI, large language model (LLM) fine-tuning, and high-throughput inference at scale, the ALT-G2-8P is positioned as an alternative to GPU-based systems for organizations building out AI infrastructure across UAE, GCC, EMEA, and APAC regions. The system ships as an integrated, rack-ready 8U or equivalent appliance and is supported by Intel's Gaudi software ecosystem, including the Intel Developer Cloud reference environment and open-source model optimizations available through Hugging Face and the Intel GitHub repositories.

Ideal for

• Large language model (LLM) pre-training and fine-tuning across multi-billion parameter architectures such as LLaMA and GPT variants
• High-throughput generative AI inference serving for enterprise NLP and content generation applications
• Computer vision model training at scale, including diffusion models and transformer-based image recognition workloads
• Distributed deep learning research requiring high aggregate HBM bandwidth and low-latency inter-accelerator communication
• Enterprise MLOps pipelines combining batch training jobs with near-real-time inference endpoints in a shared accelerator pool
• Sovereign AI infrastructure deployments in regulated industries requiring on-premises large-model compute independent of public cloud providers

Technical specifications

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