Intel Gaudi 3
AI accelerator chip using Ethernet for scaling
The Intel Gaudi 3 is an artificial intelligence accelerator designed for data centres. It uses standard Ethernet networking to connect large clusters of chips. Released in late 2024, it aims to offer a cost-effective alternative to proprietary GPU systems for running complex AI models.
- Uses standard Ethernet for cluster scaling
- Features 128GB of HBM2e memory
- Available in OAM and PCIe form factors
- Designed for large language model workloads
- Consumes 600W of power in PCIe format
What it is
The Intel Gaudi 3 is a specialised processor built for artificial intelligence workloads. It handles tasks such as training large language models and running enterprise retrieval-augmented generation systems. Intel developed this chip through its acquisition of Habana Labs. The device focuses on providing high performance while using open standards for connectivity. It allows organisations to build scalable AI infrastructure without relying on closed proprietary networks. The chip supports both matrix multiplication and vector processing engines essential for modern AI computations.
How it came about
This accelerator represents the third generation of the Gaudi product line. Intel has been developing this architecture since acquiring Habana Labs. The company announced general availability for the chip in October 2024. This launch coincided with the release of new Xeon processors. Early demonstrations occurred at industry events like Hot Chips 2024. The development aimed to create a viable competitor in the rapidly growing market for AI hardware. Intel sought to address the need for flexible and cost-effective scaling in data centres.
Inside the chip
The Gaudi 3 integrates onboard networking capabilities directly into the accelerator. It uses standard Ethernet protocols rather than proprietary interconnects. This design choice reduces the need for separate network interface cards. Each chip supports two 400GbE links externally. The memory subsystem features 128GB of HBM2e high-bandwidth memory. This configuration helps manage the massive data requirements of AI models. The architecture includes specific engines for matrix and vector operations. It supports floating-point formats such as FP8 and BF16 for efficient computation.
How it was made
The chip is available in two primary physical forms. The HL-325L is an OAM card designed for high-density server blocks. The HL-338 is a PCIe Gen5 card for standard server integration. The PCIe version operates with a thermal design power of 600 watts. It requires passive cooling and significant airflow within the chassis. Manufacturing details regarding the specific fabrication process node are not publicly specified in the provided materials. The design emphasises compatibility with existing server infrastructure.
Where it went
Organisations deploy these accelerators in large-scale data centre clusters. Systems can scale up to 1024 nodes or 8192 accelerators. Major server manufacturers like Dell and Supermicro offer compatible hardware. The Dell PowerEdge XE9680 and Supermicro X14 systems support the chip. Cloud providers such as IBM Cloud also integrate Gaudi 3 into their offerings. The hardware is suited for training and inference of multimodal AI models. It supports open-source software frameworks and tools from platforms like Hugging Face.
Versions and family
The product line includes the OAM-based HL-325L and the PCIe-based HL-338. The OAM variant fits into unified baseboards for dense configurations. The PCIe card allows integration into standard servers without specialised backplanes. Both variants share the same core computational capabilities and memory size. The predecessor to this chip was the Intel Gaudi 2. The new model offers improved input-output connectivity and compute performance compared to the previous generation.
Why it matters
Intel positions the Gaudi 3 as a cost-effective alternative to dominant GPU solutions. The company claims it offers competitive performance per dollar. By using standard Ethernet, it aims to lower the total cost of ownership for AI clusters. This approach avoids vendor lock-in associated with proprietary switching technologies. The chip supports the expansion of AI capabilities in enterprises that prefer open standards. It contributes to the diversification of the AI hardware market beyond single-supplier ecosystems.
Problems and reception
Industry observers note the strategic importance of this launch for Intel. The move to general availability marks a critical step for the company's AI ambitions. Some analysts highlight the benefit of onboard networking for reducing system complexity. Others point out that single-chip performance may not match leading competitors. However, the ability to build larger clusters efficiently is seen as a key advantage. The reliance on partner ecosystems for full stack solutions is a notable aspect of its deployment strategy.
Questions people ask about Intel Gaudi 3
What type of networking does the Gaudi 3 use?
The chip uses standard Ethernet for connectivity. It avoids proprietary technologies like NVLink. This allows users to leverage existing network infrastructure for scaling AI clusters.
How much memory does the accelerator have?
Each Gaudi 3 accelerator contains 128GB of HBM2e memory. This high-bandwidth memory is essential for handling the large datasets required by modern AI models.
Can I install this in a standard server?
Yes, the HL-338 variant comes in a PCIe Gen5 form factor. It fits into standard servers, though it requires adequate cooling and power supply due to its 600W consumption.
Who designed the Gaudi 3 chip?
Intel designed the chip following its acquisition of Habana Labs. The Gaudi line represents the third generation of technology developed by this combined team.
Good reading and sources
There is no English Wikipedia article on Intel Gaudi 3 yet, so the facts on this page come from the sources listed above: intel.com/content/www/us/en/products/details/processors/ai-accelerator, servethehome.com/intel-gaudi-3-going-ga-for-scale-out-ai-acceleration/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.