Intel Gaudi
AI accelerator series from Intel for training and inference workloads.
Intel Gaudi is a line of artificial intelligence accelerators designed for data centre tasks like large language model training. The series began with Habana Labs technology and evolved through Gaudi 2 to the Gaudi 3, which launched in April 2024. It uses standard Ethernet for networking rather than proprietary links, aiming to offer cost-effective scaling for enterprise AI deployments.
- Gaudi 3 launched in April 2024 with 128 GB HBM.
- Uses standard Ethernet for scaling up to 8,192 nodes.
- Acquired from Habana Labs by Intel in 2019.
- Available as PCIe card HL-338 or mezzanine HL-325L.
- Precedes the Falcon Shores GPU architecture.
What it is
The Intel Gaudi series consists of specialised processors built to handle heavy artificial intelligence workloads. These chips focus on training and running large models, such as those used for natural language processing. Intel acquired the original designer, Habana Labs, in late 2019 to enter this market. The latest version, Gaudi 3, arrived in April 2024. It distinguishes itself by using standard Ethernet connections for linking multiple chips together, avoiding the need for proprietary networking hardware found in some competitor systems.
How it came about
Intel purchased Habana Labs in December 2019 for two billion dollars to gain a foothold in AI acceleration. This move allowed Intel to bridge the gap between its delayed Ponte Vecchio graphics processors and the upcoming Falcon Shores architecture. The Gaudi line served as an interim solution while Intel developed its next-generation GPU plans. The company revealed pricing and performance details for Gaudi 2 and Gaudi 3 at Computex in June 2024, aiming to demonstrate value against rival products from Nvidia and AMD.
Inside the chip
Gaudi accelerators rely on matrix math units optimized for AI calculations. The Gaudi 3 model supports FP8 and BF16 precision formats, delivering high throughput for both training and inference tasks. A key design choice is the integration of standard Ethernet ports directly onto the chip. This allows clusters of up to 8,192 devices to communicate without external InfiniBand switches. The architecture supports open software frameworks, enabling developers to migrate existing models from other platforms with minimal changes to their code.
How it was made
Specific manufacturing details such as the process node or transistor count are not provided in the available records. The Gaudi 3 is available in two physical forms: a PCIe card known as HL-338 and a mezzanine card called HL-325L. The PCIe version fits into standard server slots, facilitating easy integration into existing data centre infrastructure. Intel claims this form factor simplifies deployment for enterprises looking to expand their AI capabilities without redesigning their server racks.
Where it went
These chips target data centre environments where companies train and deploy large language models. Typical applications include processing multimodal data and running enterprise retrieval-augmented generation systems. Cloud providers like IBM have integrated Gaudi 3 into their offerings, allowing users to access the hardware remotely. The accelerators are designed to handle demanding workloads that require significant memory bandwidth and computational power, serving as an alternative to traditional graphics processing units for AI tasks.
Versions and family
The family includes the earlier Gaudi 2 and the current Gaudi 3. The Gaudi 3 features 128 GB of high-bandwidth memory, compared to 80 GB in the competing Nvidia H100. Intel positions Gaudi 3 as the final iteration of this specific line before transitioning to the Falcon Shores architecture. Falcon Shores will combine Gaudi’s Ethernet fabric with Xe GPU engines. The Gaudi 3 is expected to remain in service until Falcon Shores arrives in late 2025, with a subsequent Falcon Shores 2 planned for 2026.
Why it matters
Intel aims to capture a share of the rapidly growing AI accelerator market, targeting a two billion dollar sales pipeline. By publishing list prices, Intel attempted to offer transparency in a market often dominated by opaque pricing structures. The company argues that its approach provides better cost efficiency for large-scale deployments. While sales figures trail behind market leaders like Nvidia, the Gaudi line represents Intel’s strategic effort to maintain relevance in high-performance computing and AI infrastructure.
Problems and reception
Industry analysts note that the Gaudi series offers a viable alternative for customers unable to secure supplies of dominant GPU brands. Performance comparisons suggest competitive results in specific benchmarks, particularly for large language model training. However, the ecosystem support and raw performance still lag behind the market leader. The decision to use Ethernet is viewed as a practical cost-saving measure for large clusters, though it may not match the peak bandwidth of proprietary interconnects in all scenarios.
Questions people ask about Intel Gaudi
What is the Intel Gaudi used for?
It is an AI accelerator designed for data centres. It handles training and inference for large language models and other complex artificial intelligence workloads, offering an alternative to traditional GPUs.
How does Gaudi connect multiple chips?
Unlike some competitors that use proprietary links, Gaudi uses standard Ethernet networking. This allows clusters to scale efficiently using existing infrastructure without requiring specialized switches or cables.
Who originally designed the Gaudi chips?
Habana Labs designed the initial architecture. Intel acquired Habana Labs in December 2019 for two billion dollars and has since continued the development of the Gaudi product line.
What comes after the Gaudi series?
Intel plans to replace the Gaudi line with Falcon Shores. This new architecture will merge the Ethernet fabric and matrix units from Gaudi with Xe GPU engines, with launch expected in late 2025.
Good reading and sources
There is no English Wikipedia article on Intel Gaudi yet, so the facts on this page come from the sources listed above: intel.com/content/www/us/en/products/details/processors/ai-accelerator, nextplatform.com/ai/2024/06/13/stacking-up-intel-gaudi-against-nvidia-. 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.