Intel Launches Gaudi 3 AI Accelerator to Rival Nvidia
At the Intel Vision 2024 customer and partner conference, Intel announced the Gaudi 3 AI accelerator and the Xeon 6 processor as part of its strategy to advance Enterprise AI and generative AI (GenAI) adoption. The Gaudi 3 is positioned to provide enhanced performance for AI training and inference, with claims of 40% faster training times for large language models (LLMs) compared to NVIDIA’s H100 AI chip and 50% faster inference. The accelerator boasts 1835 teraflops of FP8 compute performance, 128GB of HBM2e memory, and significant improvements over its predecessor, including double the AI FP8 power and network bandwidth, and a 50% increase in memory bandwidth.
Intel's new hardware is designed to be scalable and open, with support for an open Ethernet standard for AI workloads. The Gaudi 3 will be available in several form factors, including a mezzanine card, a universal baseboard, or a PCIe CEM specifically for the Gaudi 3. The company has established partnerships with OEMs such as Dell, Hewlett Packard Enterprise, Lenovo, and Supermicro, aiming to provide flexible and powerful AI infrastructure to enterprises.
With an industry prediction that enterprise investment in GenAI will significantly increase by 2027, Intel is targeting the challenges that enterprises face in scaling AI initiatives, such as security, integration complexities, and costs. The Gaudi 3, which is expected to be generally available in the third quarter of 2024, is designed for system scalability and will allow for clusters of AI systems with tens of thousands of accelerators connected via Ethernet.
Intel's move reflects the growing trend among tech companies to invest in AI hardware and software, aiming to meet the rising demand for infrastructure capable of deploying sophisticated AI models. As the data center AI market continues to expand, the Gaudi 3 is set to compete with other industry players like NVIDIA and AMD in the server AI chip market.
