Intel Habana Labs is the artificial intelligence accelerator business unit of Intel Corporation, formed after Intel acquired the Israeli AI chip startup Habana Labs in December 2019. The division develops specialized processors designed to accelerate machine learning and deep learning workloads, particularly for large-scale training and inference in data centers. Its product lines include the Gaudi training processors and the Greco inference processors, which compete with offerings from NVIDIA, AMD, and cloud providers' custom silicon such as AWS Trainium and Google's TPU (not in list, so omitted). The acquisition was part of Intel's broader strategy to expand beyond its traditional CPU business into the rapidly growing AI hardware market, which has been driven by the rise of large language models and generative AI systems developed by companies like OpenAI, Anthropic, and Google DeepMind.
Habana Labs was founded in 2016 by David Dahan, Ran Halutz, and Eitan Medina, all veterans of the Israeli semiconductor industry. The company was headquartered in Tel Aviv, Israel, with additional offices in San Jose, California, and other locations. Before the Intel acquisition, Habana had raised significant venture capital funding, including a $75 million Series B round in 2018 led by Intel Capital and WRV Capital, and a $120 million Series C round in November 2019, just weeks before the acquisition. Intel's acquisition was valued at approximately $2 billion, making it one of Intel's largest AI-related investments.
History and Formation
Habana Labs was established with the goal of creating purpose-built AI processors that could outperform general-purpose GPUs in both training and inference tasks. The founders had previously worked at companies such as Intel, Mellanox, and Texas Instruments, and they aimed to address the inefficiencies of using GPUs for AI workloads, which often required excessive power and memory bandwidth. The company's first product, the Goya inference processor, was announced in 2018 and demonstrated strong performance on ResNet-50 benchmarks, a common deep learning model for image classification. The Goya chip was designed to deliver high throughput with low latency, making it suitable for real-time inference applications.
In 2019, Habana introduced the Gaudi training processor, which was designed to scale efficiently for large-scale deep learning training. Gaudi featured a specialized architecture with integrated Ethernet networking, allowing multiple chips to be connected directly without the need for external switches, reducing cost and complexity in large clusters. This design was particularly aimed at data center operators who needed to train increasingly large neural networks, such as those used in natural language processing and computer vision.
Acquisition by Intel
Intel announced the acquisition of Habana Labs on December 16, 2019, for approximately $2 billion. The deal was completed in early 2020. Intel integrated Habana into its Data Platforms Group, which was later reorganized into the Accelerated Computing Systems and Graphics Group (AXG). The acquisition was part of Intel's strategy to compete in the AI accelerator market, which had been dominated by NVIDIA's GPUs. Intel had previously acquired Nervana Systems in 2016, another AI chip startup, but that project was ultimately discontinued in favor of Habana's technology. Intel's CEO at the time, Bob Swan, stated that Habana's products would complement Intel's existing portfolio and help the company address the growing demand for AI compute.
Following the acquisition, Habana continued to operate as a subsidiary, with its founders and engineers remaining in place. The division retained its Israeli headquarters and expanded its engineering teams. Intel also invested in developing software tools and frameworks to support Habana's hardware, including integration with popular machine learning frameworks like TensorFlow and PyTorch.
Products and Architecture
Intel Habana Labs' primary products are the Gaudi and Greco processors. Gaudi is designed for training large-scale deep learning models, while Greco is optimized for inference tasks. Both are based on a custom architecture that includes a mix of general-purpose CPU cores, dedicated AI compute engines, and high-bandwidth memory. The Gaudi processor, first released in 2020, features 32 GB of on-chip memory and supports up to 8,000 Gaudi processors in a single cluster via Ethernet. The second-generation Gaudi 2, launched in 2022, improved performance and added support for larger models, including those with billions of parameters.
Greco, the inference processor, was introduced in 2021 and is designed for low-latency, high-throughput inference workloads. It supports popular deep learning models such as transformers, which are the basis for large language models like GPT-4 and BERT. Greco is available in various form factors, including PCIe cards and modular systems, to fit different data center deployments.
Intel Habana Labs also provides a software stack called SynapseAI, which includes compilers, runtime libraries, and tools for optimizing models on Habana hardware. SynapseAI supports ONNX and common frameworks, allowing developers to port models with minimal changes. The software is designed to automatically optimize graph execution and memory usage, reducing the need for manual tuning.
Market Position and Competition
The AI accelerator market is highly competitive, with major players including NVIDIA, AMD, and cloud providers like Amazon Web Services (AWS) with its Trainium chips, and Google Cloud with its TPUs. Intel Habana Labs aims to differentiate itself through cost efficiency and scalability, particularly for large-scale training in data centers. Gaudi's integrated Ethernet networking is a key selling point, as it reduces the need for expensive InfiniBand switches used by NVIDIA's NVLink and InfiniBand solutions. Intel has also positioned Habana as a more open alternative, supporting industry-standard Ethernet and PCIe interfaces.
Despite these advantages, Intel Habana Labs faces challenges in gaining market share. NVIDIA's CUDA software ecosystem is deeply entrenched, and many AI developers are accustomed to using it. Intel has attempted to address this by providing compatibility layers and tools to ease migration, but adoption has been slower than initially expected. As of 2024, Intel Habana Labs' revenue contribution to Intel's overall data center business remains modest, but the division has secured several notable customers, including cloud service providers and research institutions.
Strategic Role in Intel
Intel Habana Labs is a key component of Intel's broader AI strategy, which also includes the development of the Intel Gaudi 3 processor, announced in 2024. Gaudi 3 is expected to compete directly with NVIDIA's H100 and H200 GPUs, offering comparable performance at a lower cost. Intel has also integrated Habana technology into its Xeon processors, enabling AI acceleration on general-purpose CPUs. The division works closely with Intel's foundry services, which manufacture chips for external customers, and with Intel's software teams to ensure compatibility with oneAPI, a unified programming model for heterogeneous computing.
Intel's investment in Habana reflects its belief that AI workloads will increasingly require specialized hardware, and that the company can leverage its manufacturing and packaging expertise to produce competitive AI accelerators. However, Intel has faced manufacturing challenges in recent years, including delays in its 7nm and 5nm process nodes, which have impacted the production of Gaudi chips. As of 2024, Intel has outsourced some production of Habana processors to TSMC, a major semiconductor foundry, to meet demand.
Impact and Future Outlook
Intel Habana Labs has contributed to the advancement of AI hardware by offering an alternative to GPU-based systems, particularly for organizations that prioritize cost efficiency and scalability. The division's focus on Ethernet-based networking has influenced other AI chip designers, such as Groq and SambaNova, which also use similar approaches. Intel's continued investment in Habana suggests that the company sees AI accelerators as a long-term growth area, despite the competitive challenges.
Looking ahead, Intel Habana Labs is expected to release new products that support emerging AI workloads, including large language models and generative AI. The division is also exploring ways to integrate AI acceleration into edge devices and other form factors, expanding beyond data centers. As of 2025, Intel has announced plans to spin off its foundry business, which could affect the manufacturing of Habana chips, but the division itself remains a core part of Intel's AI strategy.