Intel Gaudi 3 PCIe 96GB HBM2e – Dell PowerEdge R760

The Intel Gaudi 3 PCIe 96GB HBM2e accelerator, integrated into the Dell PowerEdge R760 platform, represents Intel's third-generation deep learning training and inference architecture purpose-built for large-scale AI workloads. The Gaudi 3 die is manufactured on TSMC's 5nm process node and delivers significant generational improvements in matrix multiplication throughput, memory bandwidth, and interconnect capacity compared to its predecessor. The accelerator features 96GB of HBM2e memory across eight HBM2e stacks, providing high-capacity, high-bandwidth memory access essential for hosting large language models and complex neural network topologies within a single device.

The PCIe form factor allows the Gaudi 3 to be deployed in standard server infrastructure without requiring proprietary fabric switches, with host connectivity provided via a PCIe Gen 5 interface. Gaudi 3 incorporates 24 100GbE RDMA network interface ports on-die, enabling direct scale-out across multiple nodes for distributed training jobs without a separate networking subsystem. The Dell PowerEdge R760 is a 2U dual-socket rack server that provides a certified, validated platform for this accelerator, offering enterprise-grade power delivery, thermal management, and systems management integration through Dell's iDRAC controller.

This solution targets enterprise data centers, cloud service operators, and AI research organizations that require scalable, standards-based AI compute infrastructure. The combination of the Gaudi 3's software ecosystem — supported through Intel's SynapseAI SDK and compatibility with PyTorch and TensorFlow via Habana community integrations — and the operational manageability of the PowerEdge R760 makes this configuration well suited for production AI deployments across sectors including financial services, healthcare analytics, telecommunications, and large-scale natural language processing initiatives.

Ideal for

• Large language model (LLM) training and fine-tuning at scale across multi-node cluster deployments using Gaudi 3's on-die RDMA networking for low-latency gradient synchronization
• Enterprise AI inference serving for transformer-based models such as BERT, GPT variants, and vision transformers requiring high memory capacity to hold full model weights on-device
• Distributed deep learning training for computer vision and recommendation system models in retail, media, and e-commerce environments
• High-throughput data center AI workloads in financial services for risk modeling, fraud detection, and algorithmic pattern recognition requiring sustained matrix compute performance
• Healthcare and life sciences AI pipelines including medical imaging analysis and genomics workloads that demand large memory footprints and reproducible, validated server hardware
• Telecommunications and edge AI research environments deploying standards-based PCIe accelerators into existing rack infrastructure without proprietary interconnect dependencies

Technical specifications

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