Gaudi 2
AI training accelerator chip by Intel
The Gaudi 2 is an artificial intelligence training processor designed by Habana Labs and sold by Intel. Announced in May 2022, it uses a 7 nanometre process to deliver higher performance than its predecessor for machine learning workloads.
- Uses TSMC 7 nanometre process technology
- Contains 24 Tensor Processor Cores
- Supports FP8 data format for AI
- Delivers 2.45 TB/sec memory bandwidth
- Consumes 650 watts of power
What it is
The Gaudi 2 is a specialised processor built for training artificial intelligence models. It belongs to the Gaudi family of chips, which originated from Habana Labs before Intel acquired the company. This device targets data centres that run heavy computational tasks like image recognition and language processing. Unlike general-purpose central processing units, it focuses on matrix mathematics required for deep learning. The chip aims to provide a cost-effective alternative to graphics processing units from competitors in the AI market.
How it came about
Intel entered the AI hardware market through acquisitions, buying Nervana Systems in 2016 and Habana Labs in 2019. These purchases aimed to secure intellectual property and talent for custom AI silicon. The first Gaudi chip launched in 2019, followed by the Gaudi 2 announcement at the Intel Vision conference in May 2022. The company positioned this release to compete with Nvidia’s GPU accelerators in the growing sector of AI compute infrastructure.
Inside the chip
The architecture relies on twenty-four Tensor Processor Cores, a significant increase from the ten found in the previous generation. These cores handle various data formats, including the new eight-bit floating point standard known as FP8. This format allows mixed precision for training and inference without converting models. The chip includes forty-eight megabytes of static random-access memory on the die. It connects to ninety-six gigabytes of high-bandwidth memory, offering substantial data throughput for large datasets.
How it was made
Taiwan Semiconductor Manufacturing Company produces the Gaudi 2 using a seven nanometre fabrication process. This shrink from the sixteen nanometre node used for the prior chip allows for more components on the same area. The manufacturing choice reflects Intel’s strategy of using external foundries for specific product lines. The increased transistor density supports the additional processing cores and memory controllers integrated into the design.
Where it went
This accelerator serves servers and workstations engaged in AI development and research. It handles tasks such as computer vision training using frameworks like TensorFlow. The chip supports natural language processing models and other compute-intensive applications in financial services and engineering. Its high bandwidth and parallel processing capabilities make it suitable for clustering in data centres where multiple cards work together on single large jobs.
Versions and family
The Gaudi 2 succeeds the Gaudi 1, also known as the HL-2000. While the earlier model featured ten tensor cores and thirty-two gigabytes of memory, the newer version expands these resources significantly. The predecessor used older memory standards and lacked support for the FP8 data type. No successor to the Gaudi 2 is mentioned in the available records, though it sits alongside other Intel AI products like Xeon processors with matrix math units.
Why it matters
The release represents Intel’s effort to capture share in the fifty billion dollar AI compute market. By offering a dedicated training engine, the company challenges the dominance of graphics processor manufacturers in this field. The move signals a shift towards specialised hardware for machine learning rather than relying solely on general-purpose CPUs. It demonstrates how major semiconductor firms are diversifying their portfolios to address specific high-growth sectors.
Questions people ask about Gaudi 2
Who designed the Gaudi 2 chip?
Habana Labs designed the Gaudi 2 before Intel acquired the company. Intel now markets and sells the chip as part of its AI accelerator portfolio, leveraging the original architecture developed by the Habana team.
What manufacturing process is used for Gaudi 2?
The chip is manufactured by Taiwan Semiconductor Manufacturing Company using a seven nanometre process. This is a shrink from the sixteen nanometre node used for the previous Gaudi 1 model, allowing for higher component density.
How much memory does the Gaudi 2 have?
It features ninety-six gigabytes of HBM2e memory across six banks. Additionally, the chip includes forty-eight megabytes of on-die static random-access memory to support fast data access for processing cores.
What is the power consumption of this accelerator?
The Gaudi 2 has a thermal design power of 650 watts. This is nearly double the three hundred and fifty watts consumed by its predecessor, reflecting the increased number of cores and higher performance capabilities.
Good reading and sources
There is no English Wikipedia article on Gaudi 2 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/2022/05/10/intel-pits-new-gaudi2-ai-training-engin. 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.