Intel Gaudi-3 AI Accelerator 128GB HBM2e 200GbE

• Brand: Intel
• Product: Intel Gaudi 3 AI Accelerator
• SKU: Gaudi-3
• Product Type: AI accelerator
• Primary Workloads: AI training, inference, fine-tuning, GenAI
• Tensor Processor Cores: 64 TPCs
• Matrix Multiplication Engines: 8 MMEs
• HBM Capacity: 128GB HBM2e
• HBM Bandwidth: 3.7TB/s
• FP8 Compute: Up to 1.8 PFLOPS
• BF16 Compute: Up to 1.8 PFLOPS
• On-die SRAM: 96MB
• Scale-Out Networking: 24 × 200GbE RoCE v2
• Bidirectional Networking Capacity: 9.6Tb/s
• Host Interface: PCIe Gen5 x16
• Architecture: Two compute dies
• Process Technology: 5nm
• Media Engine: Dedicated media processor
• Supported Frameworks: PyTorch and supported Hugging Face models
• PCIe Card Form Factor: Full-height, dual-slot, 10.5-inch for HL-338
• PCIe Card TDP: Up to 600W for HL-338
• Compatibility: Supported Intel Gaudi 3 platforms/OEM servers; PCIe Gen5 x16 for HL-338

Why Choose This Product?

• High-capacity accelerator memory — 128GB HBM2e can accommodate larger model workloads with fewer memory constraints than lower-capacity accelerator designs. 
• High memory throughput — 3.7TB/s HBM bandwidth supports data-intensive AI workloads where memory movement can influence overall accelerator utilization.
• Purpose-built AI compute — Eight MMEs and 64 TPCs provide dedicated hardware for matrix and tensor operations central to modern deep-learning workloads.
• Ethernet-native scale-out — Integrated 200GbE RoCE v2 networking enables AI cluster designs around standard Ethernet rather than requiring a proprietary accelerator interconnect. 
• Enterprise deployment flexibility — The PCIe Gen5 x16 Gaudi 3 implementation can integrate into compatible server platforms, providing an accelerator option beyond conventional GPU architectures. 
• Strong LLM positioning — Intel specifically targets Gaudi 3 at large-scale generative AI, LLM and multimodal workloads.
• Open software ecosystem — PyTorch and Hugging Face support can reduce migration friction for teams already building AI applications with these ecosystems. 
• Scale-out focused architecture — High-bandwidth integrated networking is valuable when multiple accelerators must exchange model, activation or training data efficiently.