Skip To Main Content

Intel® Gaudi® 3 AI Accelerators

High-performance AI delivers more choice, flexibility, and scale.

Intel Gaudi logo

Intel® Gaudi® 3 PCIe Card

The Intel Gaudi® 3 PCIe card (HL-338) delivers AI acceleration in a standard PCIe Gen5 form factor for seamless integration into existing servers. Designed for demanding workloads like LLMs, multimodal models, and enterprise RAG, it uses standard Ethernet to help scale efficiently and cost-effectively.

Overhead image of Intel® Gaudi® circuit board

Open

Avoid risky investments in locked, proprietary technologies such as NVLink, NVSwitch, and InfiniBand.

More I/O

Take advantage of 33 percent more I/O connectivity per accelerator compared to H1001, enabling massive scale-up and scale-out with optimized cost.

Built for Ethernet

Use the networking infrastructure you already own and support future needs with standard Ethernet hardware.

Effortless scaling

Intel Gaudi 3 AI Accelerators are designed to deliver simple, cost-effective AI scalability for the largest and most complex deployments.

Three quarter view of Gaudi AI Accelerator 32-Node Cluster

Intel Gaudi 3 AI Accelerator 32-Node Cluster reference design

Build out AI solutions with the latest Intel Gaudi 3 accelerator-based systems, made for scale and expandability with all-Ethernet-based fabrics and support for a wide range of industry AI models and frameworks.

See reference design
Exploded view of a Gaudi 3 AI accelerator

Designed for the real-world demands of AI

Intel® Gaudi® 3 AI accelerators empower you to use open, community-based software and industry-standard Ethernet networking to scale systems more flexibly. See testing results from Signal65.

Read the white paper
Intel® and Dev Zone logos

Simplify development end to end

Go from proof of concept to production in less time. From migration to deployment, Intel Gaudi 3 AI Accelerators come supported by a powerful portfolio of software tools, resources, and training. Get to know what’s available to help simplify your AI efforts. Join the developer community.

Browse developer resources

Save time and power with Intel Gaudi 3 Accelerators

Up to

2x

AI compute (FP8) vs. Intel Gaudi 2 AI Accelerators

Up to

4x

AI compute (BF16) vs. Intel Gaudi 2 AI Accelerators

Up to

2x

network bandwidth vs. Intel Gaudi 2 AI Accelerators

Deploy Intel Gaudi 3 AI Accelerators with OEMs

Develop with ease

Designed for rapid ROI. Getting started with Intel Gaudi 3 AI Accelerators is simple, regardless of whether you’re starting from scratch, fine-tuning off-the-shelf models, or migrating from a GPU-based approach.

Visit Intel Gaudi Software

Optimized for developers

Take advantage of software tools and developer resources to get up to speed effortlessly.

Integrated with PyTorch

Keep working with the library your team already knows.

Support for new and existing models

Customize reference models, start fresh, or migrate existing models using open source tools, including resources from Hugging Face.

Easy migration of GPU-based models

Quickly port your existing solutions using our purpose-built software tools.

Experience Intel Gaudi AI Accelerators in the cloud

Resources

Product and Performance Information

 

Some images on this page have been generated by AI.

 

1 NVIDIA H100 GPU (900 GB/s closed NVLink connectivity) vs. Intel® Gaudi® 3 accelerator (1200 GB/s open standard RoCE).