Update On Intel's Gaudi 3 -- January 20, 2026

 

 

As of early 2026, Intel has broadened the availability of its Gaudi 3 AI accelerators, positioning them as a cost-effective, open-standard alternative to Nvidia's H100/H200 for enterprise generative AI, training, and inference

. 

Here is the update on Intel's Gaudi 3 based on 2025/2026 developments:

1. Market Availability and Partnerships (2025-2026)

2. Performance and Technical Specifications

3. Software and Ecosystem

4. Outlook and Strategic Direction

5. Potential Challenges

• Competition: Nvidia remains the market leader. Some benchmarks in early 2025 indicated that, while Gaudi 3 is competitive, it may face stiff competition from newer H200 configurations in specific, high-end, large-scale LLM testing.
• Supply Chain: Analysts are monitoring Intel’s ability to meet 2025 volume production targets, especially given the competitive nature of the AI chip supply chain. 

IBM Cloud Adoption: IBM is a primary partner, offering Gaudi 3 as a service on IBM Cloud, including support for Red Hat OpenShift AI.

Dell AI Factory: Dell is a lead OEM incorporating Gaudi 3, particularly in PowerEdge XE9680 and XE7740 servers, which are now shipping.

Broadening Availability: Through 2025, Intel expanded availability to PCIe and rack-scale systems to support large-scale, open-standard AI clusters. 

Performance vs. Competitors: Independent testing has shown Gaudi 3 delivering competitive performance, specifically with 43% more tokens per second for small AI workloads and 36% more for large context models (e.g., Llama-3.1-405B-Instruct-FP8) compared to the Nvidia H200.

Architectural Upgrades: Gaudi 3 is manufactured using TSMC's 5nm process, delivering 4x AI compute performance for BF16 and 1.5x higher memory bandwidth (128 GB HBM2e) compared to Gaudi 2.

Networking: The chips feature 24 x 200 GbE Ethernet ports per accelerator, designed to support large-scale, open-standard Ethernet networking instead of proprietary interconnects. 

Gaudi Software 1.21.0+: Recent software updates have enhanced support for PyTorch, improved vLLM performance, and added support for models like DeepSeek-R1 and Codellama-34b.

Open Software Stack: Intel emphasizes an open-source approach, supporting tools like Intel's AI Assistant Builder, which is now on GitHub.

VMware Support: The latest VMware Cloud Foundation (9.0) supports Intel Gaudi 3, aiding in virtualized, on-premises AI deployment. 

2025/2026 Roadmap: While early 2025 saw some concerns about potential shipment target adjustments, the focus remains on leveraging Gaudi 3's cost-performance ratio for enterprises.

Future Accelerators: Gaudi 3 is part of a broader roadmap that will be succeeded by the "Falcon Shores" and "Jaguar Shores" chips in the coming years, aimed at further accelerating AI capabilities.